witcheer's picture
initial upload: turboquant KV cache benchmarks on RTX 4060 Ti 8GB
934d3cc verified
|
Raw
History Blame Contribute Delete
4.41 kB
metadata
dataset_info:
  features:
    - name: config
      dtype: string
    - name: runtime
      dtype: string
    - name: kv_cache_k
      dtype: string
    - name: kv_cache_v
      dtype: string
    - name: context_k
      dtype: int64
    - name: np_slots
      dtype: int64
    - name: graph_splits
      dtype: int64
    - name: kv_vram_mib
      dtype: int64
    - name: checkpoint_mode
      dtype: string
    - name: agent
      dtype: string
    - name: system_prompt_tokens
      dtype: int64
    - name: context_depth_k
      dtype: float64
    - name: decode_tok_s
      dtype: float64
    - name: prompt_eval_tok_s
      dtype: float64
    - name: test_type
      dtype: string
    - name: notes
      dtype: string
license: mit
tags:
  - benchmark
  - local-llm
  - 8gb-vram
  - turboquant
  - moe-offload
  - kv-cache
  - rtx-4060-ti
  - qwen
  - practitioner-tested
size_categories:
  - n<1K

RTX 4060 Ti 8GB — turboquant KV cache benchmark (Qwen3.6-35B-A3B)

practitioner-tested benchmarks of turboquant KV cache types vs standard llama.cpp on an RTX 4060 Ti 8GB with 32GB DDR5-6000 RAM.

hardware

component spec
GPU NVIDIA RTX 4060 Ti, 8 GB VRAM
CPU AMD Ryzen 5 7600X (6c/12t)
RAM 32 GB DDR5-6000 dual-channel
OS Windows 11 + WSL2 Ubuntu 26.04

model

Qwen3.6-35B-A3B-UD-Q4_K_M (22.1 GB). hybrid SSM+attention architecture — 10/40 layers use attention, 30/40 use Gated Delta Net (recurrent). MoE with partial offload: -ncmoe 30 (10 layers experts on GPU, 30 on CPU).

what this dataset covers

  1. turboquant KV cache types — turbo2, turbo3, turbo4 vs standard q4_0 and q8_0. turboquant is a llama.cpp fork (github.com/TheTom/llama-cpp-turboquant) that adds aggressive KV cache compression with auto-asymmetric behaviour (K stays q8_0 for high GQA ratios, only V is compressed).

  2. graph splits — the hidden speed determinant on 8GB VRAM. standard llama.cpp V q4_0 at 64K context produces 82 graph splits (GPU-CPU transfers per step) vs 62 at 32K. turboquant turbo2/turbo3 keep 62 splits at 64K. the split count matters more than raw VRAM usage at this boundary.

  3. context checkpoint impact — llama.cpp creates context checkpoints every 8192 tokens during prefill (~63 MiB each in VRAM). at 26K context, 5 checkpoints = 315 MiB, pushing past the 7 GB VRAM cliff. disabling with --checkpoint-every-n-tokens -1 eliminates the cliff entirely. this is a novel finding — I have not seen it documented elsewhere.

  4. agent harness overhead — Hermes Agent v0.13.0 (13.5K token system prompt) vs Pi coding agent v0.74.0 (1.9K tokens). same decode speed, 7x smaller system prompt = 13K more usable context.

  5. speed curve at depth — decode speed from 3.5K to 50K context with checkpoints disabled. smooth degradation (~2 tok/s per 10K tokens) from attention scaling on the 10 attention layers. no cliff.

key findings

  • turbo3 is the sweet spot for 8GB VRAM. same speed as turbo2 (35 tok/s) with better V cache compression. turbo4 is too aggressive (13.93 tok/s — decompression overhead kills it).
  • graph splits > raw VRAM as the speed determinant at this boundary. 62 splits = fast, 82 splits = broken.
  • context checkpoints are a hidden VRAM eater. disabling them is free on Qwen3.6 (hybrid architecture invalidates them anyway).
  • final config: 64K context, 28-34 tok/s across all depths, no cliff.

recommended config

-ngl 999 -ncmoe 30 -c 65536 -np 1 -fa on --cache-type-k q8_0 --cache-type-v turbo3 --no-cache-prompt --checkpoint-every-n-tokens -1

methodology

  • all tests on same hardware, same Q4_K_M quant, same -ncmoe 30 partial offload
  • curl tests: single 300-token completion request (warm run)
  • agent tests: real multi-step tasks via Hermes or Pi
  • decode speed from llama-server slot print_timing logs
  • tested 2026-05-11 and 2026-05-12

related datasets

collection

8GB VRAM local LLMs — practitioner-tested