# Benchmark Schema

Each line in `data/results.jsonl` is one benchmark measurement row.

JSONL files do not include a header row. The field contract below is the header.
Run `python validate.py data/results.jsonl` before submitting changes.

## Required Fields

| Field | Type | Rule |
|---|---|---|
| `variant` | string | enum: `desktop`, `sff`, `laptop` |
| `vram_gb` | number | required metadata; never identifies desktop vs SFF by itself |
| `memory_type` | string | enum: `gddr7`, `gddr6` |
| `bus_width_bit` | integer | memory bus width in bits |
| `bandwidth_gbps` | number | primary desktop/SFF variant discriminator |
| `tdp_w` | number | per-card canonical/reference total board power for the variant |
| `pcie_lanes` | integer | per-card PCIe link width for the variant, e.g. `16` for x16 |
| `gpu_count` | integer | one or more GPUs used for the run |
| `model` | string | model or run label |
| `params_b` | number | parameter count in billions |
| `quant` | string | enum below |
| `backend` | string | enum below |
| `backend_version` | string | commit, version, or digest |
| `driver_version` | string | NVIDIA driver version |
| `os` | string | OS and version |
| `context_len` | integer | prompt/context target for the run |
| `concurrency` | integer | simultaneous streams represented by the row |
| `tok_s_prefill` | number | prefill or prompt-processing tokens per second |
| `tok_s_decode` | number | decode/generation tokens per second |
| `vram_used_gb` | number | peak VRAM used by the measured run |
| `power_w` | number | board power during measured window |
| `thinking` | string | enum: `on`, `off`, `n/a` |
| `source` | string | must be `owner-measured` |
| `submitter` | string | person or organization submitting |
| `date` | string | ISO date, `YYYY-MM-DD` |
| `notes` | string | provenance and measurement caveats |

## Enums

`memory_type`:

- `gddr7`
- `gddr6`

`quant`:

- `gguf-q4_k_m`
- `gguf-q8_0`
- `nvfp4`
- `fp8`
- `bf16`
- `other`

`backend`:

- `llama.cpp`
- `vllm`
- `tensorrt-llm`
- `other`

`source`:

- `owner-measured`

## Variant Sanity Checks

The validator identifies the physical desktop/SFF variant from bandwidth, TDP,
and PCIe lanes. VRAM, memory type, and bus width are recorded metadata, but they
do not separate desktop from SFF because both are 24 GB GDDR7 ECC on a 192-bit
bus.

The validator enforces these current public specs:

| Variant | `vram_gb` | `memory_type` | `bus_width_bit` | `bandwidth_gbps` | `tdp_w` | `pcie_lanes` |
|---|---:|---|---:|---:|---:|---:|
| desktop | 24 | `gddr7` | 192 | 672 | 140 | 16 |
| sff | 24 | `gddr7` | 192 | 432 | 70 | 8 |
| laptop | 16 | `gddr7` | 256 | 896 | OEM-dependent | OEM-dependent |

For laptop rows, `tdp_w` and `pcie_lanes` still must be reported from the owner
or OEM configuration, but they are not hard-coded by the validator because mobile
implementations vary.

SFF source: NVIDIA RTX PRO 4000 Blackwell SFF Edition datasheet,
`workstation-datasheet-blackwell-rtx-pro-4000-sff-nvidia-us-4016700.pdf`.
It lists 24 GB GDDR7 ECC, 192-bit, 432 GB/s, PCIe 5.0 x8, 70 W, and dual-slot
half-height form factor.