Methods

Results are only as good as the method behind them. This section will hold the gdput benchmark methodology, workload specifications, and measurement conventions — versioned, so every published number can point at exactly how it was produced.

benchmark methodology v1 in preparation

Principles

• Prefill and decode measured separately; percentiles (p50/p95/p99) over means
• SLO set stated with every result; goodput reported alongside raw throughput
• Full configuration disclosure: hardware, engine and version, engine flags, parallelism, quantization, workload distribution — a result is a point on the config→SLO surface, and the configuration is half the datum
• Infeasible configurations (crashes, out-of-memory) recorded and published, not discarded
• Limitations and testing date on every result