# Methodology

This methodology is canonical for the benchmark dataset.

## Scope

Rows must come from owner-operated hardware. Cloud instances, vendor marketing
numbers, and third-party reposts are excluded because the value of this dataset
is reproducible local workstation behavior.

## Throughput Mode

Throughput rows should run with model "thinking" disabled when the backend and
model expose that control. Thinking can improve quality, but it changes token
budget use and makes raw tok/s comparisons noisy. Quality belongs in a separate
axis from speed.

If thinking cannot be disabled, set `thinking=n/a` and explain why in `notes`.

## Served vs Local

Record whether the model was served through `llama.cpp`, `vllm`,
`tensorrt-llm`, or another backend. Served HTTP measurements include scheduler
and loopback overhead; local CLI measurements may not. Do not merge those
without keeping the `backend` and `backend_version` fields visible.

## Clean Runs

GPU bandwidth contention invalidates results. Do not run benchmarks while another
model is using the same GPU's memory bus.

The seed rows come from llm-lab card-bench, which stops the production stack for
clean card-bench windows. Current P620 placement from DC-BRIEFING is:

- GPU 0: `gemma-aeon`, single-card production model.
- GPU 1: utility/lab card, with embed/rerank and free VRAM for suitable
  single-GPU lab work.

Practical rule for this rig: a single-GPU run that fits on the lab card does not
need a production teardown. A TP=2 or all-GPU run requires stopping
`gemma-aeon`, entering maintenance mode, running the benchmark, then restarting
and LAN-verifying the production endpoint.

For other owners: report what else was running on each GPU and how you isolated
the measurement window.

## Required Reporting

Every row must report:

- Variant: desktop, sff, or laptop.
- Bandwidth, TDP, and PCIe lanes. These separate desktop from SFF.
- VRAM, memory type, and memory bus width. These are required metadata but do
  not separate desktop from SFF because both are 24 GB GDDR7 ECC on a 192-bit
  bus.
- Driver version.
- Backend and backend version.
- Single-GPU vs multi-GPU.
- Context length and concurrency.
- Prefill tok/s and decode tok/s.
- Peak VRAM used.
- Power measurement method in `notes`.

Power should be measured from board telemetry where possible, not a wall plug
unless that is explicitly stated.