{
  "description": "Audited fields from a 120,000 request DistilBERT benchmark. Throughput and cost fields are intentionally excluded because the original harness did not measure elapsed run time.",
  "measured_hardware": {
    "accelerator": "AWS Inferentia2, inf2.xlarge",
    "gpu": "NVIDIA A10G, g5.xlarge"
  },
  "workload": {
    "model": "DistilBERT",
    "parameters": 67000000,
    "max_sequence_length": 128
  },
  "results": [
    {
      "concurrency": 1,
      "requests_each": 10000,
      "inf2_p50_ms": 2.24,
      "inf2_p99_ms": 2.41,
      "inf2_errors": 0,
      "a10g_p50_ms": 4.63,
      "a10g_p99_ms": 5.42,
      "a10g_errors": 0
    },
    {
      "concurrency": 8,
      "requests_each": 10000,
      "inf2_p50_ms": 10.47,
      "inf2_p99_ms": 17.44,
      "inf2_errors": 0,
      "a10g_p50_ms": 67.16,
      "a10g_p99_ms": 79.68,
      "a10g_errors": 0
    },
    {
      "concurrency": 32,
      "requests_each": 10000,
      "inf2_p50_ms": 44.28,
      "inf2_p99_ms": 52.87,
      "inf2_errors": 0,
      "a10g_p50_ms": 244.15,
      "a10g_p99_ms": 307.98,
      "a10g_errors": 0
    },
    {
      "concurrency": 64,
      "requests_each": 10000,
      "inf2_p50_ms": 89.24,
      "inf2_p99_ms": 99.43,
      "inf2_errors": 0,
      "a10g_p50_ms": 491.49,
      "a10g_p99_ms": 645.18,
      "a10g_errors": 0
    },
    {
      "concurrency": 128,
      "requests_each": 10000,
      "inf2_p50_ms": 178.98,
      "inf2_p99_ms": 188.92,
      "inf2_errors": 0,
      "a10g_p50_ms": 991.49,
      "a10g_p99_ms": 1188.46,
      "a10g_errors": 0
    },
    {
      "concurrency": 256,
      "requests_each": 10000,
      "inf2_p50_ms": 185.87,
      "inf2_p99_ms": 3699.71,
      "inf2_errors": 0,
      "a10g_p50_ms": 1467.65,
      "a10g_p99_ms": 5916.76,
      "a10g_errors": 29
    }
  ],
  "withdrawn_fields": [
    "throughput",
    "cost per million",
    "fleet sizing",
    "Trainium2 attribution"
  ],
  "audit_date": "2026-10-13"
}