INFERENCE LAB / BENCHMARKS

Benchmark methods

Performance claims need a workload, an environment and the raw evidence.

Build a benchmark you can audit.

Use this protocol and the downloadable recording sheet to compare serving configurations. No measured scores are presented here.

What every report will include

MeasurementReporting requirement
TTFTTime to first token, including measurement boundaries
TPOTTime per output token; report the distribution
Prefill / decode throughputSeparate input processing from output generation
VRAM usagePeak and steady-state memory, with measurement method
WorkloadConcurrency, context length, output lengths and arrival pattern
EnvironmentModel, GPU, framework, quantization, TP, versions and commands

A repeatable run sequence

  1. Freeze a workload file and record its checksum.
  2. Check outputs before measuring speed.
  3. Separate startup, warm-up and measurement windows.
  4. Run cold-prefix and warm-prefix cases separately.
  5. Repeat each configuration and record errors, latency distributions and output tokens.

Download a blank benchmark recording sheet (CSV)

Compare like with like

Record the same model, GPU, precision, tensor parallelism, context and concurrency for each comparison. Report failed requests alongside successful ones.

Read our engine evaluation method →