{
  "schema_version": 1,
  "producer": "benchmarks/sr_forward_pass/bench_forward_pass.py",
  "recorded_at": "2026-08-05T09:43:47+0000",
  "host_label": "gb10-ref",
  "harness": "benchmarks/sr_forward_pass/bench_forward_pass.py",
  "measures": "bare SPAN forward pass vs per-stage pipeline cost, batch 1",
  "scale_factor": 2,
  "weights_path": "/models/span/2xNomosUni_span_multijpg.safetensors",
  "feature_channels": 48,
  "half_precision": true,
  "device": "cuda",
  "torch_version": "2.14.0.dev20260617+cu130",
  "warmup_iters": 10,
  "timed_iters": 40,
  "canvas": "1920x1080",
  "mps_pipe_directory": "/tmp/wg3-mps",
  "mps_active_thread_pct": "50",
  "box_state_at_start": {
    "captured_at": 1785923027.0224547,
    "load_average": [
      20.77,
      36.61,
      52.94
    ],
    "cpu_count": 20,
    "mem_total_mib": 124611,
    "mem_available_mib": 56089,
    "gpu_name": "NVIDIA GB10",
    "note": "Co-tenant load on the host at this instant. Forward-pass cost can only be compared across runs taken under comparable box state."
  },
  "resolutions": [
    {
      "input_w": 640,
      "input_h": 360,
      "model_forward": {
        "n": 40,
        "mean_ms": 214.685,
        "p50_ms": 214.39,
        "p95_ms": 216.293,
        "p99_ms": 219.36,
        "min_ms": 211.912,
        "max_ms": 219.949,
        "implied_fps_at_mean": 4.66
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 236.004,
        "p50_ms": 235.758,
        "p95_ms": 241.073,
        "p99_ms": 243.604,
        "min_ms": 228.998,
        "max_ms": 254.07,
        "implied_fps_at_mean": 4.24
      },
      "sr_output_shape": "1280x720",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 19.884,
        "p50_ms": 16.639,
        "p95_ms": 41.148,
        "p99_ms": 50.635,
        "min_ms": 10.605,
        "max_ms": 52.489,
        "implied_fps_at_mean": 50.29
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 68.485,
        "p50_ms": 67.102,
        "p95_ms": 82.98,
        "p99_ms": 84.501,
        "min_ms": 56.842,
        "max_ms": 84.78,
        "implied_fps_at_mean": 14.6
      },
      "attributed": {
        "per_frame_ms": 324.373,
        "implied_fps": 3.08,
        "model_share_pct": 66.2,
        "overhead_share_pct": 33.8
      }
    },
    {
      "input_w": 854,
      "input_h": 480,
      "model_forward": {
        "n": 40,
        "mean_ms": 383.301,
        "p50_ms": 382.749,
        "p95_ms": 388.602,
        "p99_ms": 389.078,
        "min_ms": 378.712,
        "max_ms": 397.725,
        "implied_fps_at_mean": 2.61
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 555.778,
        "p50_ms": 440.136,
        "p95_ms": 695.787,
        "p99_ms": 698.369,
        "min_ms": 425.077,
        "max_ms": 700.356,
        "implied_fps_at_mean": 1.8
      },
      "sr_output_shape": "1708x960",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 21.214,
        "p50_ms": 20.808,
        "p95_ms": 27.461,
        "p99_ms": 27.832,
        "min_ms": 13.622,
        "max_ms": 28.911,
        "implied_fps_at_mean": 47.14
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 57.707,
        "p50_ms": 57.656,
        "p95_ms": 73.48,
        "p99_ms": 76.32,
        "min_ms": 42.912,
        "max_ms": 76.798,
        "implied_fps_at_mean": 17.33
      },
      "attributed": {
        "per_frame_ms": 634.699,
        "implied_fps": 1.58,
        "model_share_pct": 60.4,
        "overhead_share_pct": 39.6
      }
    },
    {
      "input_w": 1280,
      "input_h": 720,
      "model_forward": {
        "n": 40,
        "mean_ms": 899.671,
        "p50_ms": 864.099,
        "p95_ms": 932.246,
        "p99_ms": 1358.856,
        "min_ms": 852.671,
        "max_ms": 1752.167,
        "implied_fps_at_mean": 1.11
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 1237.405,
        "p50_ms": 978.435,
        "p95_ms": 1550.51,
        "p99_ms": 1552.372,
        "min_ms": 952.227,
        "max_ms": 1568.356,
        "implied_fps_at_mean": 0.81
      },
      "sr_output_shape": "2560x1440",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 12.133,
        "p50_ms": 11.247,
        "p95_ms": 16.06,
        "p99_ms": 19.604,
        "min_ms": 6.916,
        "max_ms": 20.008,
        "implied_fps_at_mean": 82.42
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 59.041,
        "p50_ms": 57.911,
        "p95_ms": 70.462,
        "p99_ms": 76.713,
        "min_ms": 46.432,
        "max_ms": 77.29,
        "implied_fps_at_mean": 16.94
      },
      "attributed": {
        "per_frame_ms": 1308.579,
        "implied_fps": 0.76,
        "model_share_pct": 68.8,
        "overhead_share_pct": 31.2
      }
    }
  ],
  "caveat": "SPAN's published 100+ fps is the authors' capability figure on their hardware. It is not a GB10 pipeline result and is not a target this project set or missed.",
  "box_state_at_end": {
    "captured_at": 1785923202.909682,
    "load_average": [
      22.89,
      29.71,
      47.32
    ],
    "cpu_count": 20,
    "mem_total_mib": 124611,
    "mem_available_mib": 56106,
    "gpu_name": "NVIDIA GB10",
    "note": "Co-tenant load on the host at this instant. Forward-pass cost can only be compared across runs taken under comparable box state."
  },
  "duration_s": 177.99,
  "arm": "quiet-mps50"
}