{
  "date": "2026-09-17",
  "hardware": {
    "cpu": "2x AMD EPYC 9334, 64 physical cores total",
    "host_memory": "approximately 1 TiB",
    "gpu": "8x Radeon Pro W7900D, approximately 48 GiB each",
    "backend": "ROCm",
    "checkpoint": "official DeepSeek V4.1 Flash safetensors"
  },
  "prefill_261933": {
    "input_tokens": 261933,
    "output_tokens": 1024,
    "recovery_baseline": {
      "ttft_s": 787.224,
      "cold_prefill_tok_s": 332.73
    },
    "retained_snapshot": {
      "ttft_s": 168.56343250721693,
      "wall_s": 232.62989360280335,
      "cold_prefill_tok_s": 1553.913539277183,
      "arena_gib_per_gpu": 3,
      "minimum_sampled_gpu_free_gib": 0.193939208984375,
      "rss_peak_gib": 196.54844284057617,
      "message_stop": true
    },
    "capacity_edge_experiment": {
      "ttft_s": 163.70356560125947,
      "cold_prefill_tok_s": 1600.044562486822,
      "arena_gib_per_gpu": 5,
      "minimum_sampled_gpu_free_gib": 0.03682708740234375,
      "retained": false,
      "reason": "insufficient VRAM safety margin"
    }
  },
  "prefill_1m": {
    "input_tokens": 1000411,
    "output_tokens": 1024,
    "ttft_s": 1365.5861574783921,
    "wall_s": 1445.3841040208936,
    "cold_prefill_tok_s": 732.587244328324,
    "minimum_sampled_gpu_free_gib": 4.175685882568359,
    "rss_peak_gib": 196.31076049804688,
    "message_stop": true,
    "output_equal_to_previous_1m_run": true
  },
  "decode": {
    "single_request_dspark_tok_s": [25.82, 26.14],
    "eight_request_aggregate_tok_s": [107.61, 109.47],
    "historical_best_eight_request_tok_s": [111.15, 113.86, 114.43]
  },
  "measurement_notes": [
    "TTFT is measured from the HTTP request until the first non-empty text event.",
    "Cold prefill throughput is input_tokens divided by TTFT and is not a pure GPU-kernel measurement.",
    "VRAM values are five-second whole-card samples and may miss instantaneous peaks.",
    "Decode and prefill rows come from different requests and must not be combined as one run."
  ]
}