{"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 1", "BW [GB/s]": "68", "GPU Cores": "7", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "108.21", "Q8_0 TG [t/s]": "7.92", "Q4_0 PP [t/s]": "107.81", "Q4_0 TG [t/s]": "14.19"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 1", "BW [GB/s]": "68", "GPU Cores": "8", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "117.25", "Q8_0 TG [t/s]": "7.91", "Q4_0 PP [t/s]": "117.96", "Q4_0 TG [t/s]": "14.15"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Pro 1", "BW [GB/s]": "200", "GPU Cores": "14", "F16 PP [t/s]": "262.65", "F16 TG [t/s]": "12.75", "Q8_0 PP [t/s]": "235.16", "Q8_0 TG [t/s]": "21.95", "Q4_0 PP [t/s]": "232.55", "Q4_0 TG [t/s]": "35.52"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Pro 1", "BW [GB/s]": "200", "GPU Cores": "16", "F16 PP [t/s]": "302.14", "F16 TG [t/s]": "12.75", "Q8_0 PP [t/s]": "270.37", "Q8_0 TG [t/s]": "22.34", "Q4_0 PP [t/s]": "266.25", "Q4_0 TG [t/s]": "36.41"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Max 1", "BW [GB/s]": "400", "GPU Cores": "24", "F16 PP [t/s]": "453.03", "F16 TG [t/s]": "22.55", "Q8_0 PP [t/s]": "405.87", "Q8_0 TG [t/s]": "37.81", "Q4_0 PP [t/s]": "400.26", "Q4_0 TG [t/s]": "54.61"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Max 1", "BW [GB/s]": "400", "GPU Cores": "32", "F16 PP [t/s]": "599.53", "F16 TG [t/s]": "23.03", "Q8_0 PP [t/s]": "537.37", "Q8_0 TG [t/s]": "40.2", "Q4_0 PP [t/s]": "530.06", "Q4_0 TG [t/s]": "61.19"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Ultra 1", "BW [GB/s]": "800", "GPU Cores": "48", "F16 PP [t/s]": "875.81", "F16 TG [t/s]": "33.92", "Q8_0 PP [t/s]": "783.45", "Q8_0 TG [t/s]": "55.69", "Q4_0 PP [t/s]": "772.24", "Q4_0 TG [t/s]": "74.93"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M1 Ultra 1", "BW [GB/s]": "800", "GPU Cores": "64", "F16 PP [t/s]": "1168.89", "F16 TG [t/s]": "37.01", "Q8_0 PP [t/s]": "1042.95", "Q8_0 TG [t/s]": "59.87", "Q4_0 PP [t/s]": "1030.04", "Q4_0 TG [t/s]": "83.73"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 2", "BW [GB/s]": "100", "GPU Cores": "8", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "147.27", "Q8_0 TG [t/s]": "12.18", "Q4_0 PP [t/s]": "145.91", "Q4_0 TG [t/s]": "21.7"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 2", "BW [GB/s]": "100", "GPU Cores": "10", "F16 PP [t/s]": "201.34", "F16 TG [t/s]": "6.72", "Q8_0 PP [t/s]": "181.4", "Q8_0 TG [t/s]": "12.21", "Q4_0 PP [t/s]": "179.57", "Q4_0 TG [t/s]": "21.91"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Pro 2", "BW [GB/s]": "200", "GPU Cores": "16", "F16 PP [t/s]": "312.65", "F16 TG [t/s]": "12.47", "Q8_0 PP [t/s]": "288.46", "Q8_0 TG [t/s]": "22.7", "Q4_0 PP [t/s]": "294.24", "Q4_0 TG [t/s]": "37.87"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Pro 2", "BW [GB/s]": "200", "GPU Cores": "19", "F16 PP [t/s]": "384.38", "F16 TG [t/s]": "13.06", "Q8_0 PP [t/s]": "344.5", "Q8_0 TG [t/s]": "23.01", "Q4_0 PP [t/s]": "341.19", "Q4_0 TG [t/s]": "38.86"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Max 2", "BW [GB/s]": "400", "GPU Cores": "30", "F16 PP [t/s]": "600.46", "F16 TG [t/s]": "24.16", "Q8_0 PP [t/s]": "540.15", "Q8_0 TG [t/s]": "39.97", "Q4_0 PP [t/s]": "537.6", "Q4_0 TG [t/s]": "60.99"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Max 2", "BW [GB/s]": "400", "GPU Cores": "38", "F16 PP [t/s]": "755.67", "F16 TG [t/s]": "24.65", "Q8_0 PP [t/s]": "677.91", "Q8_0 TG [t/s]": "41.83", "Q4_0 PP [t/s]": "671.31", "Q4_0 TG [t/s]": "65.95"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Ultra 2", "BW [GB/s]": "800", "GPU Cores": "60", "F16 PP [t/s]": "1128.59", "F16 TG [t/s]": "39.86", "Q8_0 PP [t/s]": "1003.16", "Q8_0 TG [t/s]": "62.14", "Q4_0 PP [t/s]": "1013.81", "Q4_0 TG [t/s]": "88.64"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M2 Ultra 2", "BW [GB/s]": "800", "GPU Cores": "76", "F16 PP [t/s]": "1401.85", "F16 TG [t/s]": "41.02", "Q8_0 PP [t/s]": "1248.59", "Q8_0 TG [t/s]": "66.64", "Q4_0 PP [t/s]": "1238.48", "Q4_0 TG [t/s]": "94.27"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M3 3", "BW [GB/s]": "100", "GPU Cores": "8", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟨 M3 3", "BW [GB/s]": "100", "GPU Cores": "10", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "187.52", "Q8_0 TG [t/s]": "12.27", "Q4_0 PP [t/s]": "186.75", "Q4_0 TG [t/s]": "21.34"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟨 M3 Pro 3", "BW [GB/s]": "150", "GPU Cores": "14", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "272.11", "Q8_0 TG [t/s]": "17.44", "Q4_0 PP [t/s]": "269.49", "Q4_0 TG [t/s]": "30.65"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M3 Pro 3", "BW [GB/s]": "150", "GPU Cores": "18", "F16 PP [t/s]": "357.45", "F16 TG [t/s]": "9.89", "Q8_0 PP [t/s]": "344.66", "Q8_0 TG [t/s]": "17.53", "Q4_0 PP [t/s]": "341.67", "Q4_0 TG [t/s]": "30.74"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M3 Max 3", "BW [GB/s]": "300", "GPU Cores": "30", "F16 PP [t/s]": "589.41", "F16 TG [t/s]": "19.54", "Q8_0 PP [t/s]": "566.4", "Q8_0 TG [t/s]": "34.3", "Q4_0 PP [t/s]": "567.59", "Q4_0 TG [t/s]": "56.58"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M3 Max 3", "BW [GB/s]": "400", "GPU Cores": "40", "F16 PP [t/s]": "779.17", "F16 TG [t/s]": "25.09", "Q8_0 PP [t/s]": "757.64", "Q8_0 TG [t/s]": "42.75", "Q4_0 PP [t/s]": "759.7", "Q4_0 TG [t/s]": "66.31"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M3 Ultra 3", "BW [GB/s]": "800", "GPU Cores": "60", "F16 PP [t/s]": "1121.80", "F16 TG [t/s]": "42.24", "Q8_0 PP [t/s]": "1085.76", "Q8_0 TG [t/s]": "63.55", "Q4_0 PP [t/s]": "1073.09", "Q4_0 TG [t/s]": "88.40"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M3 Ultra 3", "BW [GB/s]": "800", "GPU Cores": "80", "F16 PP [t/s]": "1538.34", "F16 TG [t/s]": "39.78", "Q8_0 PP [t/s]": "1487.51", "Q8_0 TG [t/s]": "63.93", "Q4_0 PP [t/s]": "1471.24", "Q4_0 TG [t/s]": "92.14"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M4 4", "BW [GB/s]": "120", "GPU Cores": "8", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M4 4", "BW [GB/s]": "120", "GPU Cores": "10", "F16 PP [t/s]": "230.18", "F16 TG [t/s]": "7.43", "Q8_0 PP [t/s]": "223.64", "Q8_0 TG [t/s]": "13.54", "Q4_0 PP [t/s]": "221.29", "Q4_0 TG [t/s]": "24.11"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M4 Pro 4", "BW [GB/s]": "273", "GPU Cores": "16", "F16 PP [t/s]": "381.14", "F16 TG [t/s]": "17.19", "Q8_0 PP [t/s]": "367.13", "Q8_0 TG [t/s]": "30.54", "Q4_0 PP [t/s]": "364.06", "Q4_0 TG [t/s]": "49.64"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M4 Pro 4", "BW [GB/s]": "273", "GPU Cores": "20", "F16 PP [t/s]": "464.48", "F16 TG [t/s]": "17.18", "Q8_0 PP [t/s]": "449.62", "Q8_0 TG [t/s]": "30.69", "Q4_0 PP [t/s]": "439.78", "Q4_0 TG [t/s]": "50.74"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M4 Max 4", "BW [GB/s]": "410", "GPU Cores": "32", "F16 PP [t/s]": "736.25", "F16 TG [t/s]": "24.29", "Q8_0 PP [t/s]": "718.56", "Q8_0 TG [t/s]": "43.87", "Q4_0 PP [t/s]": "713.93", "Q4_0 TG [t/s]": "69.95"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "✅ M4 Max 4", "BW [GB/s]": "546", "GPU Cores": "40", "F16 PP [t/s]": "922.83", "F16 TG [t/s]": "31.64", "Q8_0 PP [t/s]": "891.94", "Q8_0 TG [t/s]": "54.05", "Q4_0 PP [t/s]": "885.68", "Q4_0 TG [t/s]": "83.06"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M4 Ultra", "BW [GB/s]": "820", "GPU Cores": "64", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M4 Ultra", "BW [GB/s]": "1092", "GPU Cores": "80", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 5", "BW [GB/s]": "154", "GPU Cores": "8", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 5", "BW [GB/s]": "154", "GPU Cores": "10", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Pro 5", "BW [GB/s]": "307", "GPU Cores": "16", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Pro 5", "BW [GB/s]": "307", "GPU Cores": "20", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Max 5", "BW [GB/s]": "460", "GPU Cores": "32", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Max 5", "BW [GB/s]": "614", "GPU Cores": "40", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Ultra", "BW [GB/s]": "?", "GPU Cores": "?", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "🟥 M5 Ultra", "BW [GB/s]": "?", "GPU Cores": "?", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "", "BW [GB/s]": "BW [GB/s]", "GPU Cores": "GPU Cores", "F16 PP [t/s]": "F16 PP [t/s]", "F16 TG [t/s]": "F16 TG [t/s]", "Q8_0 PP [t/s]": "Q8_0 PP [t/s]", "Q8_0 TG [t/s]": "Q8_0 TG [t/s]", "Q4_0 PP [t/s]": "Q4_0 PP [t/s]", "Q4_0 TG [t/s]": "Q4_0 TG [t/s]"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "2023 Nov 21", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "M2 Ultra 8e672ef", "BW [GB/s]": "800", "GPU Cores": "76", "F16 PP [t/s]": "1401.85", "F16 TG [t/s]": "41.02", "Q8_0 PP [t/s]": "1248.59", "Q8_0 TG [t/s]": "66.64", "Q4_0 PP [t/s]": "1238.48", "Q4_0 TG [t/s]": "94.27"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "2024 Nov 12", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "M2 Ultra 86ed72d + FA", "BW [GB/s]": "800", "GPU Cores": "76", "F16 PP [t/s]": "1525.95", "F16 TG [t/s]": "43.15", "Q8_0 PP [t/s]": "1368.18", "Q8_0 TG [t/s]": "73.11", "Q4_0 PP [t/s]": "1391.78", "Q4_0 TG [t/s]": "108.80"} {"source": "apple-silicon", "Summary\nLLaMA 7B": "2025 Aug 02", "BW [GB/s]": "", "GPU Cores": "", "F16 PP [t/s]": "", "F16 TG [t/s]": "", "Q8_0 PP [t/s]": "", "Q8_0 TG [t/s]": "", "Q4_0 PP [t/s]": "", "Q4_0 TG [t/s]": ""} {"source": "apple-silicon", "Summary\nLLaMA 7B": "M2 Ultra 5c0eb5e + FA", "BW [GB/s]": "800", "GPU Cores": "76", "F16 PP [t/s]": "1561.35", "F16 TG [t/s]": "43.24", "Q8_0 PP [t/s]": "1386.97", "Q8_0 TG [t/s]": "73.35", "Q4_0 PP [t/s]": "1412.42", "Q4_0 TG [t/s]": "109.41"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5090", "Memory": "32 GB / GDDR7 / 512 bit", "pp512 t/s": "14073.41 ± 115.16", "tg128 t/s": "290.02 ± 1.10", "Commit": "8cf6b42", "Thanks to": "@totaldev"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX PRO 6000 Blackwell", "Memory": "96 GB / GDDR7 / 512 bit", "pp512 t/s": "14854.63 ± 22.73", "tg128 t/s": "274.20 ± 0.14", "Commit": "79c1160", "Thanks to": "@Tom94"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "H100 80 GB", "Memory": "80 GB / HBM3 / 5120 bit", "pp512 t/s": "9918.34 ± 176.97", "tg128 t/s": "267.81 ± 1.54", "Commit": "5143fa8", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A100 80 GB", "Memory": "80 GB / HBM2e / 5120 bit", "pp512 t/s": "4849.53 ± 8.94", "tg128 t/s": "190.88 ± 0.33", "Commit": "5143fa8", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4090 D", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "10293.86 ± 134.72", "tg128 t/s": "189.33 ± 0.19", "Commit": "79c1160", "Thanks to": "@autonomous-AI-lab"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4090", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "11992.70 ± 107.99", "tg128 t/s": "186.21 ± 0.13", "Commit": "2241453", "Thanks to": "@lhl"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5080", "Memory": "16 GB / GDDR7 / 256 bit", "pp512 t/s": "8297.36 ± 9.50", "tg128 t/s": "181.99 ± 0.42", "Commit": "8a4280c", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5070 Ti", "Memory": "16 GB / GDDR7 / 256 bit", "pp512 t/s": "6952.38 ± 13.73", "tg128 t/s": "176.85 ± 0.07", "Commit": "933414c", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 6000 Ada", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "9229.23 ± 101.78", "tg128 t/s": "176.07 ± 0.26", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3090 Ti", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "6567.49 ± 20.30", "tg128 t/s": "171.19 ± 3.98", "Commit": "9c35706", "Thanks to": "@slaren"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3090", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "5174.69 ± 21.83", "tg128 t/s": "158.16 ± 0.21", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "L40", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "8870.49 ± 378.76", "tg128 t/s": "152.01 ± 0.28", "Commit": "ee09828", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4080 SUPER", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "8125.15 ± 41.05", "tg128 t/s": "148.33 ± 0.20", "Commit": "81086cd", "Thanks to": "@zacharyarnaise"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4080", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "8031.64 ± 26.49", "tg128 t/s": "142.49 ± 0.16", "Commit": "20638e4", "Thanks to": "@Ristovski"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3080", "Memory": "10 GB / GDDR6X / 320 bit", "pp512 t/s": "5013.86 ± 24.80", "tg128 t/s": "139.65 ± 0.99", "Commit": "9c35706", "Thanks to": "@slaren"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A6000", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "4913.93 ± 6.79", "tg128 t/s": "138.73 ± 2.75", "Commit": "4795c91", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4070 Ti SUPER", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "6924.53 ± 13.87", "tg128 t/s": "132.26 ± 0.16", "Commit": "9c35706", "Thanks to": "@Ristovski"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX PRO 4000 Blackwell", "Memory": "24 GB / GDDR7 / 192 bit", "pp512 t/s": "4992.83 ± 113.52", "tg128 t/s": "131.66 ± 0.20", "Commit": "7d77f07", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A5000", "Memory": "24 GB / GDDR6 / 384 bit", "pp512 t/s": "4028.16 ± 19.14", "tg128 t/s": "130.07 ± 2.74", "Commit": "e5155e6", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla V100", "Memory": "32 GB / HBM2 / 4096 bit", "pp512 t/s": "3042.64 ± 40.71", "tg128 t/s": "129.08 ± 0.05", "Commit": "51f5a45", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5070", "Memory": "12 GB / GDDR7 / 192 bit", "pp512 t/s": "5184.75 ± 18.70", "tg128 t/s": "127.54 ± 0.46", "Commit": "", "Thanks to": "@Spyro000"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A40", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "4609.01 ± 10.67", "tg128 t/s": "124.11 ± 0.17", "Commit": "3470a5c", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A30", "Memory": "24 GB / HBM2e / 3072 bit", "pp512 t/s": "2767.10 ± 1.88", "tg128 t/s": "124.81 ± 0.16", "Commit": "583cb83", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Titan V", "Memory": "12 GB / HBM2 / 3072 bit", "pp512 t/s": "2617.46 ± 2.10", "tg128 t/s": "108.79 ± 0.05", "Commit": "e56abd2", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2080 Ti", "Memory": "11 GB / GDDR6 / 352 bit", "pp512 t/s": "2890.66 ± 2.42", "tg128 t/s": "107.51 ± 0.21", "Commit": "9c35706", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 6000", "Memory": "24 GB / GDDR6 / 384 bit", "pp512 t/s": "2751.18 ± 19.43", "tg128 t/s": "102.77 ± 0.04", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 8000", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "2709.95 ± 3.35", "tg128 t/s": "102.68 ± 0.03", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4500", "Memory": "20 GB / GDDR6 / 320 bit", "pp512 t/s": "2827.20 ± 66.43", "tg128 t/s": "97.32 ± 2.80", "Commit": "5cdb27e", "Thanks to": "@aleksyx"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5060 Ti", "Memory": "16 GB / GDDR7 / 128 bit", "pp512 t/s": "3737.25 ± 6.79", "tg128 t/s": "90.94 ± 0.02", "Commit": "89d1029", "Thanks to": "@mike-llamacpp"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2070 SUPER", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "2088.34 ± 1.94", "tg128 t/s": "88.06 ± 0.28", "Commit": "bc07349", "Thanks to": "@phstudy"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4000", "Memory": "16 GB / GDDR6 / 256 bit", "pp512 t/s": "2684.06 ± 15.28", "tg128 t/s": "83.77 ± 0.37", "Commit": "65349f2", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Titan Xp", "Memory": "12 GB / GDDR5X / 384 bit", "pp512 t/s": "1154.96 ± 1.46", "tg128 t/s": "76.08 ± 0.08", "Commit": "c4510dc", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3060", "Memory": "12 GB / GDDR6 / 192 bit", "pp512 t/s": "2137.50 ± 10.12", "tg128 t/s": "75.57 ± 0.07", "Commit": "baa9255", "Thanks to": "@QuantiusBenignus"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 4000", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "1536.89 ± 0.90", "tg128 t/s": "65.62 ± 0.62", "Commit": "7d77f07", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4060 Ti", "Memory": "8 GB / GDDR6 / 128 bit", "pp512 t/s": "3394.63 ± 7.44", "tg128 t/s": "63.86 ± 0.01", "Commit": "89d1029", "Thanks to": "@mike-llamacpp"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1080 Ti", "Memory": "11 GB / GDDR5X / 352 bit", "pp512 t/s": "1084.41 ± 3.01", "tg128 t/s": "62.49 ± 0.06", "Commit": "9c35706", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4000 Ada", "Memory": "20 GB / GDDR6 / 160 bit", "pp512 t/s": "2779.77 ± 9.91", "tg128 t/s": "61.83 ± 0.04", "Commit": "a74a0d6", "Thanks to": "@sdwolfz"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2060 SUPER", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "1420.24 ± 1.95", "tg128 t/s": "60.04 ± 0.01", "Commit": "5c0eb5e", "Thanks to": "@ggerganov"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P100", "Memory": "16 GB / HBM2 / 4096 bit", "pp512 t/s": "760.80 ± 2.92", "tg128 t/s": "58.35 ± 0.00", "Commit": "b8372ee", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "DGX Spark", "Memory": "128 GB / LPDDR5x", "pp512 t/s": "3062.31 ± 11.02", "tg128 t/s": "57.21 ± 0.06", "Commit": "5acd455", "Thanks to": "@ggerganov"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P40", "Memory": "24 GB / GDDR5 / 384 bit", "pp512 t/s": "1007.42 ± 1.23", "tg128 t/s": "54.74 ± 0.07", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2000 Ada", "Memory": "16 GB / GDDR6 / 128 bit", "pp512 t/s": "1956.22 ± 7.74", "tg128 t/s": "50.62 ± 0.04", "Commit": "756cfea", "Thanks to": "@DigitalRudeness"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla T4", "Memory": "16 GB / GDDR6 / 256 bit", "pp512 t/s": "1219.06 ± 4.18", "tg128 t/s": "46.38 ± 0.73", "Commit": "d32e03f", "Thanks to": "@pt13762104"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4050 Laptop", "Memory": "6 GB / GDDR6 / 96 bit", "pp512 t/s": "1725.85 + 17.85", "tg128 t/s": "43.72 + 0.41", "Commit": "d79d8f3", "Thanks to": "@TimCabbage"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1660", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "148.91 ± 0.01", "tg128 t/s": "41.35 ± 0.02", "Commit": "9515c61", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla M40", "Memory": "24 GB / GDDR5 / 384 bit", "pp512 t/s": "282.65 ± 0.15", "tg128 t/s": "38.04 ± 0.02", "Commit": "97d5117", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1070 Ti", "Memory": "8 GB / GDDR5 / 256 bit", "pp512 t/s": "714.44 ± 2.04", "tg128 t/s": "37.82 ± 0.02", "Commit": "79c1160", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Jetson AGX Orin", "Memory": "64 GB / LPDDR5 / 256 bit", "pp512 t/s": "991.31 ± 1.15", "tg128 t/s": "33.58 ± 0.14", "Commit": "c1b1876", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P4", "Memory": "8 GB / GDDR5 / 256 bit", "pp512 t/s": "514.53 ± 3.06", "tg128 t/s": "33.29 ± 0.00", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "P106-100", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "406.94 ± 0.25", "tg128 t/s": "30.40 ± 0.02", "Commit": "5fd160b", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1060", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "416.85 ± 1.75", "tg128 t/s": "27.79 ± 0.02", "Commit": "5fd160b", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro T1000", "Memory": "4 GB / GDDR5 / 128 bit", "pp512 t/s": "79.44 ± 0.01", "tg128 t/s": "27.82 ± 0.18", "Commit": "f6da8cb", "Thanks to": "@hanabu"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro P2000", "Memory": "5 GB / GDDR5 / 160 bit", "pp512 t/s": "309.30 ± 0.05", "tg128 t/s": "23.63 ± 0.00", "Commit": "baa9255", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro P1000", "Memory": "4 GB / GDDR5 / 128 bit", "pp512 t/s": "183.40 ± 0.11", "tg128 t/s": "13.99 ± 0.13", "Commit": "1e74897", "Thanks to": "@aleksyx"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla K80", "Memory": "12 GB / GDDR5 / 384 bit", "pp512 t/s": "133.14 ± 0.55", "tg128 t/s": "13.80 ± 0.02", "Commit": "32732f2", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Chip", "Memory": "Memory", "pp512 t/s": "pp512 t/s", "tg128 t/s": "tg128 t/s", "Commit": "Commit", "Thanks to": "Thanks to"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5090", "Memory": "32 GB / GDDR7 / 512 bit", "pp512 t/s": "14970.15 ± 381.06", "tg128 t/s": "300.40 ± 0.28", "Commit": "8cf6b42", "Thanks to": "@totaldev"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX PRO 6000 Blackwell", "Memory": "96 GB / GDDR7 / 512 bit", "pp512 t/s": "16618.98 ± 20.66", "tg128 t/s": "281.11 ± 0.41", "Commit": "5143fa8", "Thanks to": "@Tom94"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "H100 80 GB", "Memory": "80 GB / HBM3 / 5120 bit", "pp512 t/s": "11263.29 ± 98.34", "tg128 t/s": "280.74 ± 1.17", "Commit": "5143fa8", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A100 80 GB", "Memory": "80 GB / HBM2e / 5120 bit", "pp512 t/s": "5285.96 ± 6.58", "tg128 t/s": "200.90 ± 0.12", "Commit": "5143fa8", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4090 D", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "12506.97 ± 11.51", "tg128 t/s": "191.57 ± 0.03", "Commit": "79c1160", "Thanks to": "@autonomous-AI-lab"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4090", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "14770.63 ± 102.93", "tg128 t/s": "188.96 ± 0.05", "Commit": "2241453", "Thanks to": "@lhl"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5080", "Memory": "16 GB / GDDR7 / 256 bit", "pp512 t/s": "9487.70 ± 21.89", "tg128 t/s": "184.68 ± 0.05", "Commit": "8a4280c", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5070 Ti", "Memory": "16 GB / GDDR7 / 256 bit", "pp512 t/s": "8419.56 ± 35.50", "tg128 t/s": "182.43 ± 0.09", "Commit": "933414c", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 6000 Ada", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "10576.85 ± 530.21", "tg128 t/s": "179.47 ± 0.32", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3090 Ti", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "6924.01 ± 10.76", "tg128 t/s": "172.26 ± 1.31", "Commit": "9c35706", "Thanks to": "@slaren"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX PRO 4500 Blackwell", "Memory": "32 GB / GDDR7 / 256 bit", "pp512 t/s": "7251.66 ± 92.40", "tg128 t/s": "168.90 ± 0.20", "Commit": "becc481", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3090", "Memory": "24 GB / GDDR6X / 384 bit", "pp512 t/s": "5560.06 ± 16.28", "tg128 t/s": "161.89 ± 0.18", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "L40", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "10097.64 ± 671.22", "tg128 t/s": "153.76 ± 0.12", "Commit": "ee09828", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4080 SUPER", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "9439.01 ± 56.75", "tg128 t/s": "147.48 ± 1.41", "Commit": "81086cd", "Thanks to": "@zacharyarnaise"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4080", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "9205.93 ± 22.31", "tg128 t/s": "143.47 ± 0.02", "Commit": "20638e4", "Thanks to": "@Ristovski"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A6000", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "5662.39 ± 13.87", "tg128 t/s": "144.87 ± 0.18", "Commit": "4795c91", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3080", "Memory": "10 GB / GDDR6X / 320 bit", "pp512 t/s": "5569.56 ± 14.04", "tg128 t/s": "139.95 ± 0.95", "Commit": "9c35706", "Thanks to": "@slaren"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX PRO 4000 Blackwell", "Memory": "24 GB / GDDR7 / 192 bit", "pp512 t/s": "5674.44 ± 139.53", "tg128 t/s": "136.38 ± 0.13", "Commit": "7d77f07", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A5000", "Memory": "24 GB / GDDR6 / 384 bit", "pp512 t/s": "4552.15 ± 9.68", "tg128 t/s": "135.83 ± 0.11", "Commit": "e5155e6", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla V100", "Memory": "32 GB / HBM2 / 4096 bit", "pp512 t/s": "2973.78 ± 3.62", "tg128 t/s": "134.76 ± 0.02", "Commit": "51f5a45", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4070 Ti SUPER", "Memory": "16 GB / GDDR6X / 256 bit", "pp512 t/s": "7612.32 ± 37.35", "tg128 t/s": "132.85 ± 0.31", "Commit": "9c35706", "Thanks to": "@Ristovski"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5070", "Memory": "12 GB / GDDR7 / 192 bit", "pp512 t/s": "5783.44 ± 36.95", "tg128 t/s": "128.21 ± 2.52", "Commit": "", "Thanks to": "@Spyro000"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A40", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "5256.38 ± 19.39", "tg128 t/s": "126.24 ± 0.06", "Commit": "3470a5c", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "A30", "Memory": "24 GB / HBM2e / 3072 bit", "pp512 t/s": "3068.72 ± 0.63", "tg128 t/s": "131.93 ± 0.18", "Commit": "583cb83", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Titan V", "Memory": "12 GB / HBM2 / 3072 bit", "pp512 t/s": "2481.25 ± 1.31", "tg128 t/s": "112.17 ± 0.01", "Commit": "e56abd2", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2080 Ti", "Memory": "11 GB / GDDR6 / 352 bit", "pp512 t/s": "3107.61 ± 4.34", "tg128 t/s": "109.17 ± 0.07", "Commit": "9c35706", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 6000", "Memory": "24 GB / GDDR6 / 384 bit", "pp512 t/s": "3053.96 ± 1.37", "tg128 t/s": "104.38 ± 0.04", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 8000", "Memory": "48 GB / GDDR6 / 384 bit", "pp512 t/s": "3052.35 ± 5.64", "tg128 t/s": "103.63 ± 0.02", "Commit": "b8e09f0", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4500", "Memory": "20 GB / GDDR6 / 320 bit", "pp512 t/s": "3453.10 ± 49.19", "tg128 t/s": "103.00 ± 0.25", "Commit": "5cdb27e", "Thanks to": "@aleksyx"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 5060 Ti", "Memory": "16 GB / GDDR7 / 128 bit", "pp512 t/s": "4195.53 ± 1.98", "tg128 t/s": "93.46 ± 0.01", "Commit": "89d1029", "Thanks to": "@mike-llamacpp"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2070 SUPER", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "2293.29 ± 5.91", "tg128 t/s": "87.71 ± 0.29", "Commit": "bc07349", "Thanks to": "@phstudy"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4000", "Memory": "16 GB / GDDR6 / 256 bit", "pp512 t/s": "2807.83 ± 52.44", "tg128 t/s": "85.17 ± 0.66", "Commit": "65349f2", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 3060", "Memory": "12 GB / GDDR6 / 192 bit", "pp512 t/s": "2407.67 ± 3.73", "tg128 t/s": "76.92 ± 0.03", "Commit": "baa9255", "Thanks to": "@QuantiusBenignus"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Titan Xp", "Memory": "12 GB / GDDR5X / 384 bit", "pp512 t/s": "1218.12 ± 1.82", "tg128 t/s": "73.84 ± 0.04", "Commit": "c4510dc", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro RTX 4000", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "1662.80 ± 2.04", "tg128 t/s": "67.62 ± 0.67", "Commit": "7d77f07", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 4060 Ti", "Memory": "8 GB / GDDR6 / 128 bit", "pp512 t/s": "3803.45 ± 70.80", "tg128 t/s": "64.03 ± 0.53", "Commit": "89d1029", "Thanks to": "@mike-llamacpp"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX A4000 Ada", "Memory": "20 GB / GDDR6 / 160 bit", "pp512 t/s": "3171.86 ± 4.34", "tg128 t/s": "61.37 ± 0.01", "Commit": "a74a0d6", "Thanks to": "@sdwolfz"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P100", "Memory": "16 GB / HBM2 / 4096 bit", "pp512 t/s": "787.36 ± 3.27", "tg128 t/s": "61.99 ± 0.00", "Commit": "b8372ee", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1080 Ti", "Memory": "11 GB / GDDR5X / 352 bit", "pp512 t/s": "1138.14 ± 2.02", "tg128 t/s": "61.38 ± 0.03", "Commit": "9c35706", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2060 SUPER", "Memory": "8 GB / GDDR6 / 256 bit", "pp512 t/s": "1563.77 ± 0.51", "tg128 t/s": "61.13 ± 0.05", "Commit": "5c0eb5e", "Thanks to": "@ggerganov"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "DGX Spark", "Memory": "128 GB / LPDDR5x", "pp512 t/s": "3661.37 ± 38.66", "tg128 t/s": "56.74 ± 0.03", "Commit": "5acd455", "Thanks to": "@ggerganov"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P40", "Memory": "24 GB / GDDR5 / 384 bit", "pp512 t/s": "1079.66 ± 0.18", "tg128 t/s": "53.73 ± 0.05", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "RTX 2000 Ada", "Memory": "16 GB / GDDR6 / 128 bit", "pp512 t/s": "2250.14 ± 5.91", "tg128 t/s": "50.71 ± 0.01", "Commit": "756cfea", "Thanks to": "@DigitalRudeness"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla T4", "Memory": "16 GB / GDDR6 / 256 bit", "pp512 t/s": "1309.73 ± 1.02", "tg128 t/s": "44.03 ± 0.57", "Commit": "d32e03f", "Thanks to": "@pt13762104"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1660", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "154.45 ± 0.52", "tg128 t/s": "41.43 ± 0.01", "Commit": "9515c61", "Thanks to": "@ariya"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla M40", "Memory": "24 GB / GDDR5 / 384 bit", "pp512 t/s": "290.17 ± 0.11", "tg128 t/s": "39.98 ± 0.01", "Commit": "97d5117", "Thanks to": "@Hedede"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1070 Ti", "Memory": "8 GB / GDDR5 / 256 bit", "pp512 t/s": "790.52 ± 2.39", "tg128 t/s": "37.87 ± 0.00", "Commit": "79c1160", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Jetson AGX Orin", "Memory": "64 GB / LPDDR5 / 256 bit", "pp512 t/s": "1171.96 ± 4.70", "tg128 t/s": "35.88 ± 0.18", "Commit": "c1b1876", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla P4", "Memory": "8 GB / GDDR5 / 256 bit", "pp512 t/s": "529.53 ± 2.12", "tg128 t/s": "33.12 ± 0.03", "Commit": "c76b420", "Thanks to": "@m18coppola"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "P106-100", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "438.49 ± 0.38", "tg128 t/s": "30.64 ± 0.06", "Commit": "5fd160b", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "GTX 1060", "Memory": "6 GB / GDDR5 / 192 bit", "pp512 t/s": "446.19 ± 0.81", "tg128 t/s": "28.18 ± 0.01", "Commit": "5fd160b", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro T1000", "Memory": "4 GB / GDDR5 / 128 bit", "pp512 t/s": "27.46 ± 0.23", "tg128 t/s": "27.46 ± 0.23", "Commit": "f6da8cb", "Thanks to": "@hanabu"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro P2000", "Memory": "5 GB / GDDR5 / 160 bit", "pp512 t/s": "311.55 ± 0.19", "tg128 t/s": "23.76 ± 0.01", "Commit": "baa9255", "Thanks to": "@TinyServal"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Tesla K80", "Memory": "12 GB / GDDR5 / 384 bit", "pp512 t/s": "133.36 ± 0.60", "tg128 t/s": "14.27 ± 0.32", "Commit": "32732f2", "Thanks to": "@pebaryan"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "Quadro P1000", "Memory": "4 GB / GDDR5 / 128 bit", "pp512 t/s": "173.82 ± 0.02", "tg128 t/s": "13.65 ± 0.14", "Commit": "1e74897", "Thanks to": "@aleksyx"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "it's sad that nsys and ncu can't be used on these cards\n==ERROR== ERR_NVCMPGPU - Profiling is not supported on the NVIDIA Crypto Mining Processors (CMP) of the target device 0. For more information, please visit https://developer.nvidia.com/ERR_NVCMPGPU\n\nbut to add more color to the above results:\nGPU utilization during inference — nvidia-smi dmon -s u while running the above benchmarks\n\n\n\nRegime", "Memory": "SM active % (avg)", "pp512 t/s": "SM active % (max)", "tg128 t/s": "HBM2e controller % (avg)", "Commit": "HBM2e controller % (max)", "Thanks to": "Bottleneck"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "pp512 (prompt processing)", "Memory": "96", "pp512 t/s": "100", "tg128 t/s": "3", "Commit": "5", "Thanks to": "compute (SM)"} {"source": "cuda", "This is similar to the Performance of llama.cpp on Apple Silicon M-series, Performance of llama.cpp on AMD ROCm(HIP) and Performance of llama.cpp with Vulkan, but for CUDA! I think it's good to consolidate and discuss our results here.\nWe'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our CUDA releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\nllama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0,1\n\nShare your llama-bench results along with the git hash and CUDA info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same device I'll prioritize newer commits with substantial CUDA updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that your memory speed and number of channels will greatly affect your inference speed!\nCUDA Scoreboard for Llama 2 7B, Q4_0 (no FA)\n\n\n\nChip": "tg128 (token generation)", "Memory": "95", "pp512 t/s": "100", "tg128 t/s": "14", "Commit": "18", "Thanks to": "compute (SM)"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5090", "pp512 t/s": "10381.64 ± 508.84", "tg128 t/s": "263.63 ± 0.91", "Commit": "ca71fb9", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 XTX", "pp512 t/s": "3531.93 ± 31.74", "tg128 t/s": "191.28 ± 0.20", "Commit": "2f0c2db", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4090", "pp512 t/s": "9452.03 ± 187.70", "tg128 t/s": "187.97 ± 0.21", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5080", "pp512 t/s": "7444.99 ± 20.11", "tg128 t/s": "185.10 ± 0.54", "Commit": "f6b533d", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3090", "pp512 t/s": "4666.15 ± 12.89", "tg128 t/s": "164.05 ± 0.89", "Commit": "d05fe1d", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia A100", "pp512 t/s": "6389.86 ± 4.83", "tg128 t/s": "160.78 ± 0.16", "Commit": "2257758", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4080 Super", "pp512 t/s": "7101.18 ± 269.79", "tg128 t/s": "147.13 ± 5.64", "Commit": "81086cd", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3080", "pp512 t/s": "4287.11 ± 55.50", "tg128 t/s": "139.15 ± 0.05", "Commit": "7c7d6ce", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A5000", "pp512 t/s": "3641.55 ± 9.05", "tg128 t/s": "139.89 ± 0.69", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9070 XT", "pp512 t/s": "5036.04 ± 88.16", "tg128 t/s": "137.11 ± 0.02", "Commit": "e9fd8dc", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5070 Ti", "pp512 t/s": "6213.63 ± 27.72", "tg128 t/s": "135.63 ± 0.18", "Commit": "d13d0f6", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon AI Pro R9700", "pp512 t/s": "5281.22 ± 48.93", "tg128 t/s": "133.27 ± 0.33", "Commit": "f9f3365", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla V100", "pp512 t/s": "1391.39 ± 1.19", "tg128 t/s": "129.58 ± 0.58", "Commit": "7d77f07", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Ti Super", "pp512 t/s": "6099.18 ± 154.30", "tg128 t/s": "129.45 ± 0.18", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 XT", "pp512 t/s": "2941.58 ± 17.17", "tg128 t/s": "123.18 ± 0.40", "Commit": "71e74a3", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9070", "pp512 t/s": "3164.10 ± 66.84", "tg128 t/s": "119.71 ± 3.40", "Commit": "21c17b5", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7800 XT", "pp512 t/s": "2017.33 ± 19.30", "tg128 t/s": "118.27 ± 0.27", "Commit": "4fdbc1e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 GRE", "pp512 t/s": "2336.31 ± 7.52", "tg128 t/s": "116.11 ± 0.26", "Commit": "4b2a477", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M3 Ultra", "pp512 t/s": "1116.83 ± 0.55", "tg128 t/s": "115.54 ± 0.78", "Commit": "2d451c8", "Comments": "MoltenVK"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B70", "pp512 t/s": "3379.00 ± 47.92", "tg128 t/s": "112.02 ± 1.08", "Commit": "b863507", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia A30", "pp512 t/s": "3290.70 ± 12.03", "tg128 t/s": "111.48 ± 0.18", "Commit": "583cb83", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Ti", "pp512 t/s": "4981.44 ± 102.35", "tg128 t/s": "110.53 ± 0.00", "Commit": "516a4ca", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Super", "pp512 t/s": "4608.20 ± 31.66", "tg128 t/s": "108.74 ± 0.18", "Commit": "c945aaa", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Instinct MI50", "pp512 t/s": "1119.55 ± 1.50", "tg128 t/s": "108.51 ± 0.81", "Commit": "3af34b9", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6900 XT", "pp512 t/s": "1901.20 ± 36.70", "tg128 t/s": "108.00 ± 0.03", "Commit": "a972fae", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro VII", "pp512 t/s": "912.47 ± 1.06", "tg128 t/s": "106.03 ± 0.89", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Titan V", "pp512 t/s": "796.29 ± 5.84", "tg128 t/s": "105.06 ± 0.27", "Commit": "e56abd2", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon VII", "pp512 t/s": "1059.14 ± 0.56", "tg128 t/s": "101.19 ± 0.53", "Commit": "77d6ae4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800 XT", "pp512 t/s": "1752.92 ± 1.71", "tg128 t/s": "100.32 ± 0.97", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 2080 Ti", "pp512 t/s": "1888.24 ± 9.20", "tg128 t/s": "97.58 ± 6.60", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800", "pp512 t/s": "1698.69 ± 0.80", "tg128 t/s": "95.61 ± 0.19", "Commit": "4b385bf", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W6800X Duo", "pp512 t/s": "687.71 ± 4.33", "tg128 t/s": "94.82 ± 0.12", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5060 Ti", "pp512 t/s": "3460.92 ± 7.16", "tg128 t/s": "93.51 ± 0.15", "Commit": "89f10ba", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070", "pp512 t/s": "3179.37 ± 46.16", "tg128 t/s": "92.29 ± 0.28", "Commit": "9a48399", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3060 Ti", "pp512 t/s": "2361.43 ± 4.32", "tg128 t/s": "86.74 ± 0.34", "Commit": "b42c7fa", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W6800X", "pp512 t/s": "510.80 ± 0.13", "tg128 t/s": "86.47 ± 0.46", "Commit": "13b4548", "Comments": "MoltenVK"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6700 XT", "pp512 t/s": "1051.20 ± 0.98", "tg128 t/s": "83.88 ± 0.08", "Commit": "6d75883", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6750 XT", "pp512 t/s": "1040.58 ± 0.35", "tg128 t/s": "81.98 ± 0.03", "Commit": "228f34c", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro V620", "pp512 t/s": "1595.32 ± 1.59", "tg128 t/s": "81.78 ± 0.06", "Commit": "03d4698", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3070", "pp512 t/s": "2113.02 ± 7.38", "tg128 t/s": "78.71 ± 0.13", "Commit": "1b8fb81", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Instinct MI60", "pp512 t/s": "369.26 ± 2.48", "tg128 t/s": "78.16 ± 1.40", "Commit": "504af20", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3060", "pp512 t/s": "1815.70 ± 5.85", "tg128 t/s": "75.94 ± 0.80", "Commit": "92c0b38", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M4 Max", "pp512 t/s": "724.77 ± 20.93", "tg128 t/s": "75.02 ± 0.14", "Commit": "1ece0cb6", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla T10", "pp512 t/s": "1692.70 ± 2.05", "tg128 t/s": "75.01 ± 0.21", "Commit": "7f76692", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A4000", "pp512 t/s": "2248.14 ± 7.59", "tg128 t/s": "73.74 ± 0.08", "Commit": "f5245b5", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 5700 XT", "pp512 t/s": "529.69 ± 0.26", "tg128 t/s": "70.73 ± 0.04", "Commit": "4fdbc1e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9060 XT", "pp512 t/s": "2141.67 ± 6.87", "tg128 t/s": "70.54 ± 0.74", "Commit": "ed52f36", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc B580", "pp512 t/s": "620.94 ± 15.33", "tg128 t/s": "70.14 ± 0.28", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro V540", "pp512 t/s": "583.88 ± 6.56", "tg128 t/s": "69.64 ± 0.24", "Commit": "9da3dcd", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W5700", "pp512 t/s": "449.85 ± 0.46", "tg128 t/s": "68.55 ± 0.15", "Commit": "23bc779", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B60", "pp512 t/s": "522.36 ± 3.60", "tg128 t/s": "68.55 ± 0.01", "Commit": "516a4ca", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1080 Ti", "pp512 t/s": "540.69 ± 0.71", "tg128 t/s": "64.99 ± 0.08", "Commit": "360d653", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 2070 Super", "pp512 t/s": "1199.13 ± 7.70", "tg128 t/s": "64.64 ± 0.20", "Commit": "b7552cf", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3070 Mobile", "pp512 t/s": "1689.40 ± 19.57", "tg128 t/s": "63.64 ± 0.39", "Commit": "ceff6bb", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX Vega 64", "pp512 t/s": "575.06 ± 1.03", "tg128 t/s": "63.43 ± 0.07", "Commit": "5dcb711", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P100", "pp512 t/s": "678.14 ± 1.40", "tg128 t/s": "63.16 ± 0.06", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD BC-250", "pp512 t/s": "370.66 ± 0.04", "tg128 t/s": "62.32 ± 0.32", "Commit": "5886f4f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6650 XT", "pp512 t/s": "1029.52 ± 1.21", "tg128 t/s": "62.14 ± 0.02", "Commit": "dbb852b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4060 Mobile", "pp512 t/s": "2135.66 ± 23.18", "tg128 t/s": "59.53 ± 0.03", "Commit": "a5c07dc", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P40", "pp512 t/s": "488.06 ± 0.27", "tg128 t/s": "59.36 ± 0.16", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1660 Ti Mobile", "pp512 t/s": "511.67 ± 2.85", "tg128 t/s": "56.60 ± 0.07", "Commit": "b43556e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Instinct MI25", "pp512 t/s": "439.42 ± 0.34", "tg128 t/s": "54.69 ± 0.03", "Commit": "2739a71", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6600 XT", "pp512 t/s": "574.65 ± 0.86", "tg128 t/s": "53.92 ± 0.11", "Commit": "091592d", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen Al Max+ 395", "pp512 t/s": "1288.96 ± 6.49", "tg128 t/s": "53.59 ± 0.38", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7600 XT", "pp512 t/s": "840.85 ± 3.02", "tg128 t/s": "53.02 ± 0.01", "Commit": "01d8eaa", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A770", "pp512 t/s": "1073.85 + 29.68", "tg128 t/s": "52.56 + 0.11", "Commit": "a69d54f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GB10", "pp512 t/s": "2737.79 ± 19.56", "tg128 t/s": "52.28 ± 0.03", "Commit": "b9da444", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro S9300 x2", "pp512 t/s": "247.26 ± 0.43", "tg128 t/s": "51.86 ± 0.11", "Commit": "eec1e33", "Comments": "Split across two GPUs"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6600", "pp512 t/s": "761.89 ± 1.76", "tg128 t/s": "50.63 ± 0.02", "Commit": "b1c70e2", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX Vega 56", "pp512 t/s": "439.87 ± 0.61", "tg128 t/s": "50.23 ± 0.14", "Commit": "92c0b38", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc B570", "pp512 t/s": "913.95 ± 0.90", "tg128 t/s": "49.64 ± 0.03", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3060 Mobile", "pp512 t/s": "1059.76 ± 3.54", "tg128 t/s": "49.03 ± 0.13", "Commit": "dbb3a47", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800M", "pp512 t/s": "861.99 ± 7.67", "tg128 t/s": "48.71 ± 0.71", "Commit": "8e6f8bc", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6600M", "pp512 t/s": "605.59 ± 0.65", "tg128 t/s": "48.21 ± 0.07", "Commit": "fe5b78c", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A770M", "pp512 t/s": "875.92 ± 2.16", "tg128 t/s": "47.69 ± 0.16", "Commit": "eeee367", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B50", "pp512 t/s": "1186.45 ± 4.43", "tg128 t/s": "46.30 ± 0.07", "Commit": "aa46bda", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia P104-100", "pp512 t/s": "311.90 ± 0.22", "tg128 t/s": "46.18 ± 0.05", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A2000", "pp512 t/s": "1245.19 ± 8.76", "tg128 t/s": "45.52 ± 0.54", "Commit": "b1afcab", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7600M XT", "pp512 t/s": "459.39 ± 2.34", "tg128 t/s": "45.28 ± 0.10", "Commit": "b9ab0a4", "Comments": "eGPU"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro V340", "pp512 t/s": "375.41 ± 0.24", "tg128 t/s": "45.16 ± 0.06", "Commit": "9da3dcd", "Comments": "Split across two GPUs"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1070 Ti", "pp512 t/s": "297.50 ± 0.54", "tg128 t/s": "42.86 ± 1.20", "Commit": "860a9e4", "Comments": "eGPU"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A750", "pp512 t/s": "1075.94 ± 13.89", "tg128 t/s": "42.66 ± 0.18", "Commit": "c1b1876", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4050 Mobile", "pp512 t/s": "1154.28 + 15.76", "tg128 t/s": "41.89 + 0.10", "Commit": "d79d8f3", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1070", "pp512 t/s": "321.57 ± 0.93", "tg128 t/s": "41.48 ± 0.09", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla M40", "pp512 t/s": "92.48 ± 0.02", "tg128 t/s": "39.35 ± 1.22", "Commit": "b8372ee", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 580", "pp512 t/s": "258.03 ± 0.71", "tg128 t/s": "39.32 ± 0.03", "Commit": "de4c07f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 470", "pp512 t/s": "218.07 ± 0.56", "tg128 t/s": "38.63 ± 0.21", "Commit": "e288693", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 570", "pp512 t/s": "226.11 ± 0.45", "tg128 t/s": "37.44 ± 0.03", "Commit": "e583f3b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W5500", "pp512 t/s": "315.39 ± 3.76", "tg128 t/s": "36.82 ± 0.38", "Commit": "860a9e4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 480", "pp512 t/s": "248.66 ± 0.28", "tg128 t/s": "34.71 ± 0.14", "Commit": "3b15924", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2 Ultra", "pp512 t/s": "205.98 ± 0.02", "tg128 t/s": "34.34 ± 0.12", "Commit": "dbb852b", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 980", "pp512 t/s": "186.24 ± 0.09", "tg128 t/s": "33.90 ± 0.51", "Commit": "860a9e4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia P106-100", "pp512 t/s": "183.78 ± 0.26", "tg128 t/s": "29.77 ± 0.04", "Commit": "23bc779", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro W8100", "pp512 t/s": "155.22 ± 0.17", "tg128 t/s": "29.52 ± 0.05", "Commit": "4536363", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P4", "pp512 t/s": "265.54 ± 0.21", "tg128 t/s": "28.03 ± 0.14", "Commit": "24d2ee0", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6500 XT", "pp512 t/s": "255.25 ± 0.35", "tg128 t/s": "27.81 ± 0.10", "Commit": "g9fdfcd", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M3", "pp512 t/s": "263.70 ± 0.02", "tg128 t/s": "26.39 ± 0.14", "Commit": "b9ab0a4", "Comments": "MoltenVK"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro S10000", "pp512 t/s": "94.78 ± 0.02", "tg128 t/s": "25.32 ± 0.02", "Commit": "914a82d", "Comments": "Split across two GPUs"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Quadro P2000", "pp512 t/s": "169.55 ± 0.17", "tg128 t/s": "23.05 ± 0.03", "Commit": "63f8fe0", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core Ultra 200 Series", "pp512 t/s": "544.95 ± 4.15", "tg128 t/s": "22.49 ± 0.09", "Commit": "cea560f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen AI 9 300 Series", "pp512 t/s": "479.07 ± 0.41", "tg128 t/s": "22.41 ± 0.18", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 6000 Series", "pp512 t/s": "240.89 ± 0.52", "tg128 t/s": "21.26 ± 0.08", "Commit": "ee09828", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2 Pro", "pp512 t/s": "62.70 ± 0.03", "tg128 t/s": "20.95 ± 0.11", "Commit": "1fe0029", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1050 Ti", "pp512 t/s": "136.42 ± 0.67", "tg128 t/s": "20.96 ± 0.21", "Commit": "2f0c2db", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 7000 Series", "pp512 t/s": "281.62 ± 1.56", "tg128 t/s": "19.91 ± 0.07", "Commit": "ebce03e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 8000 Series", "pp512 t/s": "267.27 ± 7.61", "tg128 t/s": "19.81 ± 0.08", "Commit": "1e9d771", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen Z1 Extreme", "pp512 t/s": "199.36 ± 7.02", "tg128 t/s": "18.77 ± 0.02", "Commit": "53ff6b9", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro D700", "pp512 t/s": "69.95 ± 0.04", "tg128 t/s": "16.62 ± 0.01", "Commit": "d3bd719", "Comments": "MoltenVK, running in FP16 mode on FP32 only chip"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro WX 4100", "pp512 t/s": "78.79 ± 0.10", "tg128 t/s": "16.05 ± 0.07", "Commit": "860a9e4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2", "pp512 t/s": "50.79 ± 0.16", "tg128 t/s": "13.50 ± 0.02", "Commit": "8c0d6bb", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M1", "pp512 t/s": "38.29 ± 0.00", "tg128 t/s": "12.47 ± 0.03", "Commit": "2370665", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 5000 Series", "pp512 t/s": "90.55 ± 0.08", "tg128 t/s": "10.98 ± 0.07", "Commit": "d84635b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core 1100 Series", "pp512 t/s": "187.20 ± 1.78", "tg128 t/s": "10.39 ± 0.04", "Commit": "abb9f3c", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 550", "pp512 t/s": "52.66 ± 0.49", "tg128 t/s": "10.20 ± 0.01", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 4000 Series", "pp512 t/s": "103.87 ± 0.02", "tg128 t/s": "9.63 ± 0.01", "Commit": "4b385bf", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla K80", "pp512 t/s": "89.46 ± 0.10", "tg128 t/s": "9.39 ± 0.06", "Commit": "5d46bab", "Comments": "Running on single GPU"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla K40", "pp512 t/s": "64.37 ± 0.09", "tg128 t/s": "9.30 ± 0.19", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "MediaTek Dimensity 9400", "pp512 t/s": "38.36 ± 15.15", "tg128 t/s": "8.92 ± 0.06", "Commit": "b9ab0a4", "Comments": "GPU supports coopmat but pp512 is faster with it turned off"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core Ultra 100 Series", "pp512 t/s": "185.51 ± 0.22", "tg128 t/s": "8.21 ± 0.07", "Commit": "1d72c84", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 3000 Series", "pp512 t/s": "48.63 ± 0.10", "tg128 t/s": "8.49 ± 0.01", "Commit": "1fe0029", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core 1000 Series", "pp512 t/s": "25.58 ± 0.00", "tg128 t/s": "4.25 ± 0.18", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core 8000 Series", "pp512 t/s": "25.43 ± 0.17", "tg128 t/s": "3.35 ± 0.03", "Commit": "c4df49a", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel N150", "pp512 t/s": "28.84 ± 0.02", "tg128 t/s": "2.93 ± 0.00", "Commit": "4f63cd7", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "CIX CD8180", "pp512 t/s": "2.80 ± 0.01", "tg128 t/s": "5.51 ± 0.00", "Commit": "4dca015", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Chip", "pp512 t/s": "pp512 t/s", "tg128 t/s": "tg128 t/s", "Commit": "Commit", "Comments": "Comments"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5090", "pp512 t/s": "11796.38 ± 601.36", "tg128 t/s": "273.68 ± 0.52", "Commit": "ca71fb9", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 XTX", "pp512 t/s": "3332.90 ± 11.47", "tg128 t/s": "195.30 ± 0.23", "Commit": "2f0c2db", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5080", "pp512 t/s": "8054.59 ± 35.68", "tg128 t/s": "192.17 ± 0.21", "Commit": "f6b533d", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4090", "pp512 t/s": "10830.41 ± 36.25", "tg128 t/s": "190.10 ± 0.31", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5070 Ti", "pp512 t/s": "7567.94 ± 41.96", "tg128 t/s": "176.18 ± 1.12", "Commit": "0d0764d", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3090", "pp512 t/s": "5200.68 ± 16.44", "tg128 t/s": "171.61 ± 2.69", "Commit": "d05fe1d", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia A100", "pp512 t/s": "7064.40 ± 1.63", "tg128 t/s": "170.56 ± 0.02", "Commit": "2257758", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4080 Super", "pp512 t/s": "8007.37 ± 46.03", "tg128 t/s": "150.20 ± 0.26", "Commit": "81086cd", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3080", "pp512 t/s": "4913.83 ± 21.52", "tg128 t/s": "145.74 ± 0.16", "Commit": "7c7d6ce", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla V100", "pp512 t/s": "1411.25 ± 2.12", "tg128 t/s": "142.13 ± 0.03", "Commit": "7d77f07", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A5000", "pp512 t/s": "4071.22 ± 13.13", "tg128 t/s": "140.43 ± 0.22", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Ti Super", "pp512 t/s": "6801.18 ± 40.12", "tg128 t/s": "135.81 ± 4.29", "Commit": "4ae88d0", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon AI Pro R9700", "pp512 t/s": "5620.02 ± 60.27", "tg128 t/s": "138.79 ± 0.10", "Commit": "f9f3365", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9070 XT", "pp512 t/s": "5048.07 ± 2.12", "tg128 t/s": "131.54 ± 0.05", "Commit": "e9fd8dc", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7800 XT", "pp512 t/s": "2197.05 ± 6.03", "tg128 t/s": "124.86 ± 0.10", "Commit": "4fdbc1e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 XT", "pp512 t/s": "2701.13 ± 8.75", "tg128 t/s": "120.62 ± 0.36", "Commit": "71e74a3", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9070", "pp512 t/s": "2859.98 ± 31.53", "tg128 t/s": "119.51 ± 0.13", "Commit": "21c17b5", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Instinct MI50", "pp512 t/s": "1127.37 ± 0.48", "tg128 t/s": "117.94 ± 0.02", "Commit": "3af34b9", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia A30", "pp512 t/s": "3158.25 ± 22.59", "tg128 t/s": "117.43 ± 0.23", "Commit": "583cb83", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B70", "pp512 t/s": "3150.55 ± 9.06", "tg128 t/s": "114.19 ± 1.07", "Commit": "b863507", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Ti", "pp512 t/s": "5515.26 ± 7.92", "tg128 t/s": "113.52 ± 0.00", "Commit": "516a4ca", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070 Super", "pp512 t/s": "5198.91 ± 29.02", "tg128 t/s": "112.00 ± 0.17", "Commit": "c945aaa", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon VII", "pp512 t/s": "1007.19 ± 6.55", "tg128 t/s": "110.86 ± 0.34", "Commit": "77d6ae4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7900 GRE", "pp512 t/s": "2251.37 ± 2.91", "tg128 t/s": "110.50 ± 0.23", "Commit": "4b2a477", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Titan V", "pp512 t/s": "792.74 ± 4.30", "tg128 t/s": "109.21 ± 0.72", "Commit": "e56abd2", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro VII", "pp512 t/s": "783.94 ± 0.77", "tg128 t/s": "108.45 ± 0.48", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6900 XT", "pp512 t/s": "1761.93 ± 4.75", "tg128 t/s": "106.15 ± 0.04", "Commit": "a972fae", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 2080 Ti", "pp512 t/s": "1936.25 ± 32.08", "tg128 t/s": "100.99 ± 0.24", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800 XT", "pp512 t/s": "1704.79 ± 0.71", "tg128 t/s": "100.50 ± 0.06", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W6800X Duo", "pp512 t/s": "795.28 ± 0.72", "tg128 t/s": "100.08 ± 0.02", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 5060 Ti", "pp512 t/s": "3912.65 ± 5.86", "tg128 t/s": "97.01 ± 0.14", "Commit": "89f10ba", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800", "pp512 t/s": "1749.46 ± 3.36", "tg128 t/s": "96.65 ± 0.48", "Commit": "4b385bf", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4070", "pp512 t/s": "4293.57 ± 27.70", "tg128 t/s": "91.49 ± 0.89", "Commit": "9a48399", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3060 Ti", "pp512 t/s": "2644.95 ± 26.82", "tg128 t/s": "90.35 ± 1.00", "Commit": "b42c7fa", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6750 XT", "pp512 t/s": "997.05 ± 0.45", "tg128 t/s": "82.29 ± 0.06", "Commit": "228f34c", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6700 XT", "pp512 t/s": "1010.90 ± 12.89", "tg128 t/s": "81.86 ± 0.19", "Commit": "6d75883", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3060", "pp512 t/s": "2012.88 ± 10.12", "tg128 t/s": "80.59 ± 0.02", "Commit": "92c0b38", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro V620", "pp512 t/s": "1556.31 ± 2.82", "tg128 t/s": "79.24 ± 0.09", "Commit": "03d4698", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A4000", "pp512 t/s": "2482.74 ± 26.05", "tg128 t/s": "76.07 ± 0.08", "Commit": "f5245b5", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla T10", "pp512 t/s": "1840.14 ± 1.22", "tg128 t/s": "76.05 ± 0.13", "Commit": "7f76692", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 5700 XT", "pp512 t/s": "538.31 ± 0.35", "tg128 t/s": "74.43 ± 0.03", "Commit": "4fdbc1e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc B580", "pp512 t/s": "419.49 ± 3.37", "tg128 t/s": "72.00 ± 0.24", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M4 Max", "pp512 t/s": "557.46 ± 26.87", "tg128 t/s": "71.79 ± 4.16", "Commit": "1ece0cb6", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro W5700", "pp512 t/s": "446.98 ± 0.39", "tg128 t/s": "71.30 ± 0.24", "Commit": "23bc779", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B60", "pp512 t/s": "274.76 ± 0.27", "tg128 t/s": "70.54 ± 0.03", "Commit": "516a4ca", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 9060 XT", "pp512 t/s": "1915.41 ± 7.90", "tg128 t/s": "70.52 ± 0.16", "Commit": "ed52f36", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX Vega 64", "pp512 t/s": "584.98 ± 1.12", "tg128 t/s": "67.70 ± 0.09", "Commit": "5dcb711", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P100", "pp512 t/s": "685.51 ± 0.88", "tg128 t/s": "66.48 ± 0.02", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1080 Ti", "pp512 t/s": "529.96 ± 0.38", "tg128 t/s": "64.63 ± 0.10", "Commit": "360d653", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD BC-250", "pp512 t/s": "356.87 ± 1.24", "tg128 t/s": "63.14 ± 0.09", "Commit": "5886f4f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 3070 Mobile", "pp512 t/s": "1832.07 ± 57.14", "tg128 t/s": "62.92 ± 0.37", "Commit": "ceff6bb", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6650 XT", "pp512 t/s": "1088.90 ± 0.40", "tg128 t/s": "64.53 ± 0.75", "Commit": "dbb852b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX 4060 Mobile", "pp512 t/s": "2358.03 ± 12.17", "tg128 t/s": "60.01 ± 0.08", "Commit": "a5c07dc", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P40", "pp512 t/s": "484.37 ± 0.27", "tg128 t/s": "59.22 ± 0.15", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1660 Ti Mobile", "pp512 t/s": "514.34 ± 0.88", "tg128 t/s": "57.30 ± 0.42", "Commit": "b43556e", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 7600 XT", "pp512 t/s": "1024.38 ± 7.56", "tg128 t/s": "56.11 ± 0.02", "Commit": "01d8eaa", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro S9300 x2", "pp512 t/s": "243.33 ± 0.22", "tg128 t/s": "55.64 ± 0.06", "Commit": "eec1e33", "Comments": "Split across two GPUs"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GB10", "pp512 t/s": "3279.89 ± 26.78", "tg128 t/s": "53.64 ± 0.05", "Commit": "b9da444", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6600", "pp512 t/s": "808.76 ± 0.15", "tg128 t/s": "53.24 ± 0.03", "Commit": "b1c70e2", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A770", "pp512 t/s": "1119.68 + 30.25", "tg128 t/s": "53.07 + 0.09", "Commit": "a69d54f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen Al Max+ 395", "pp512 t/s": "1357.07 ± 10.94", "tg128 t/s": "53.00 ± 0.13", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX Vega 56", "pp512 t/s": "428.54 ± 0.50", "tg128 t/s": "52.66 ± 0.03", "Commit": "92c0b38", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc B570", "pp512 t/s": "288.51 ± 0.09", "tg128 t/s": "50.49 ± 0.05", "Commit": "7f76692", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 6800M", "pp512 t/s": "784.16 ± 2.76", "tg128 t/s": "49.06 ± 0.34", "Commit": "8e6f8bc", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia RTX A2000", "pp512 t/s": "1361.85 ± 3.26", "tg128 t/s": "45.69 ± 0.20", "Commit": "b1afcab", "Comments": "coopmat2"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A770M", "pp512 t/s": "384.74 ± 0.78", "tg128 t/s": "45.68 ± 0.06", "Commit": "eeee367", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia P104-100", "pp512 t/s": "325.30 ± 0.25", "tg128 t/s": "48.64 ± 0.04", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc Pro B50", "pp512 t/s": "1122.53 ± 3.73", "tg128 t/s": "47.69 ± 0.06", "Commit": "aa46bda", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro V340", "pp512 t/s": "360.23 ± 0.74", "tg128 t/s": "47.54 ± 0.06", "Commit": "9da3dcd", "Comments": "Split across two GPUs"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Arc A750", "pp512 t/s": "303.37 ± 1.44", "tg128 t/s": "43.96 ± 0.03", "Commit": "c1b1876", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1070 Ti", "pp512 t/s": "292.85 ± 0.23", "tg128 t/s": "43.42 ± 0.34", "Commit": "860a9e4", "Comments": "eGPU"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1070", "pp512 t/s": "330.84 ± 1.02", "tg128 t/s": "43.33 ± 0.06", "Commit": "360d653", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla M40", "pp512 t/s": "93.35 ± 0.01", "tg128 t/s": "41.68 ± 0.01", "Commit": "b8372ee", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 570", "pp512 t/s": "229.23 ± 0.45", "tg128 t/s": "40.18 ± 0.00", "Commit": "e583f3b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 470", "pp512 t/s": "197.26 ± 0.27", "tg128 t/s": "37.28 ± 0.11", "Commit": "3769fe6", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon RX 480", "pp512 t/s": "194.52 ± 0.61", "tg128 t/s": "37.23 ± 0.09", "Commit": "0bcb40b", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2 Ultra", "pp512 t/s": "198.83 ± 0.85", "tg128 t/s": "198.83 ± 0.85", "Commit": "dbb852b", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 980", "pp512 t/s": "180.97 ± 0.74", "tg128 t/s": "34.16 ± 0.10", "Commit": "860a9e4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia P106-100", "pp512 t/s": "183.40 ± 0.34", "tg128 t/s": "30.79 ± 0.32", "Commit": "23bc779", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD FirePro W8100", "pp512 t/s": "140.52 ± 0.34", "tg128 t/s": "29.28 ± 0.14", "Commit": "4536363", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla P4", "pp512 t/s": "287.14 ± 0.29", "tg128 t/s": "28.37 ± 0.24", "Commit": "24d2ee0", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Quadro P2000", "pp512 t/s": "181.71 ± 0.12", "tg128 t/s": "23.77 ± 0.02", "Commit": "63f8fe0", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core Ultra 200 Series", "pp512 t/s": "536.48 ± 1.27", "tg128 t/s": "23.05 ± 0.04", "Commit": "cea560f", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen AI 9 300 Series", "pp512 t/s": "532.59 ± 3.55", "tg128 t/s": "22.31 ± 0.06", "Commit": "N/A", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 8000 Series", "pp512 t/s": "311.02 ± 0.12", "tg128 t/s": "21.28 ± 0.01", "Commit": "1e9d771", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 6000 Series", "pp512 t/s": "277.91 ± 0.37", "tg128 t/s": "21.15 ± 0.09", "Commit": "ee09828", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2 Pro", "pp512 t/s": "58.86 ± 0.02", "tg128 t/s": "20.97 ± 0.03", "Commit": "1fe0029", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 7000 Series", "pp512 t/s": "312.85 ± 2.51", "tg128 t/s": "20.09 ± 0.35", "Commit": "835b2b9", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia GTX 1050 Ti", "pp512 t/s": "127.54 ± 1.03", "tg128 t/s": "20.08 ± 0.17", "Commit": "2f0c2db", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Radeon Pro WX 4100", "pp512 t/s": "75.59 ± 0.19", "tg128 t/s": "16.56 ± 0.04", "Commit": "860a9e4", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M1", "pp512 t/s": "35.93 ± 0.00", "tg128 t/s": "12.85 ± 0.02", "Commit": "2370665", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Apple M2", "pp512 t/s": "46.81 ± 0.08", "tg128 t/s": "12.25 ± 2.30", "Commit": "8c0d6bb", "Comments": "Asahi Linux"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 5000 Series", "pp512 t/s": "79.06 ± 0.01", "tg128 t/s": "10.75 ± 0.00", "Commit": "5d195f1", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core 1100 Series", "pp512 t/s": "174.77 ± 4.47", "tg128 t/s": "10.58 ± 0.03", "Commit": "abb9f3c", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla K40", "pp512 t/s": "64.37 ± 0.02", "tg128 t/s": "9.92 ± 0.06", "Commit": "eec1e33", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 4000 Series", "pp512 t/s": "113.32 ± 0.01", "tg128 t/s": "9.87 ± 0.01", "Commit": "4b385bf", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Nvidia Tesla K80", "pp512 t/s": "88.26 ± 0.19", "tg128 t/s": "9.49 ± 0.01", "Commit": "5d46bab", "Comments": "Running on single GPU"} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "AMD Ryzen 5 3000 Series", "pp512 t/s": "47.41 ± 0.14", "tg128 t/s": "8.47 ± 0.01", "Commit": "1fe0029", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core Ultra 100 Series", "pp512 t/s": "77.66 ± 2.75", "tg128 t/s": "7.75 ± 0.05", "Commit": "2e89f76", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel Core 8000 Series", "pp512 t/s": "25.55 ± 0.04", "tg128 t/s": "3.35 ± 0.02", "Commit": "c4df49a", "Comments": ""} {"source": "vulkan", "This is similar to the Apple Silicon benchmark thread, but for Vulkan! We'll be testing the Llama 2 7B model like the other thread to keep things consistent, and use Q4_0 as it's simple to compute and small enough to fit on a 4GB GPU. You can download it here.\nInstructions\nEither run the commands below or download one of our Vulkan releases. If you have multiple GPUs please run the test on a single GPU using -sm none -mg YOUR_GPU_NUMBER unless the model is too big to fit in VRAM. If you use RADV please run with the environment variable RADV_PERFTEST=nogttspill as that can fix a bunch of performance issues.\nwget https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf\ngit clone https://github.com/ggerganov/llama.cpp\ncd llama.cpp\nmkdir build\ncd build\ncmake .. -DGGML_VULKAN=on -DCMAKE_BUILD_TYPE=Release\nmake\n./bin/llama-bench -m ../../llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1 (add any extra options here)\n\nShare your llama-bench results along with the git hash and Vulkan info string in the comments. Feel free to try other models and compare backends, but only valid runs will be placed on the scoreboard.\nIf multiple entries are posted for the same setup I'll prioritize newer commits with substantial Vulkan updates, otherwise I'll pick the one with the highest overall score at my discretion. Performance may vary depending on driver, operating system, board manufacturer, etc. even if the chip is the same. For integrated graphics note that the memory speed and number of channels will greatly affect your inference speed!\nVulkan Scoreboard (Click on the headings to expand the section)\n\nLlama 2 7B, Q4_0, no FA\n\n\n\nChip": "Intel N150", "pp512 t/s": "25.59 ± 0.00", "tg128 t/s": "2.91 ± 0.00", "Commit": "4f63cd7", "Comments": ""}