Benchmarks

SLO-aware measurements of real inference workloads. Every result publishes the model, hardware, engine and version, quantization, context length, concurrency, SLO set, TTFT, ITL, throughput, goodput, cost, limitations, and testing date — as an article and as raw JSON/CSV.

first benchmark dataset in preparation

Structured result format

{
  "model": "example-moe-model",
  "hardware": "8x NVIDIA H200",
  "engine": "vLLM",
  "quantization": "FP8",
  "context_length": 80000,
  "concurrency": 16,
  "ttft_p95_ms": 4200,
  "itl_p95_ms": 48,
  "slo_compliant_requests_percent": 91.4,
  "effective_output_tokens_per_second": 730,
  "tested_at": "2026-07-21",
  "methodology": "https://gdput.com/methods"
}