{
  "schema_version": 1,
  "producer": "benchmarks/sr_forward_pass/bench_forward_pass.py",
  "recorded_at": "2026-08-05T09:39:51+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": null,
  "mps_active_thread_pct": null,
  "box_state_at_start": {
    "captured_at": 1785922791.2825453,
    "load_average": [
      34.37,
      53.24,
      61.48
    ],
    "cpu_count": 20,
    "mem_total_mib": 124611,
    "mem_available_mib": 55585,
    "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": 124.063,
        "p50_ms": 123.889,
        "p95_ms": 125.839,
        "p99_ms": 127.028,
        "min_ms": 122.179,
        "max_ms": 127.072,
        "implied_fps_at_mean": 8.06
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 144.031,
        "p50_ms": 143.951,
        "p95_ms": 146.671,
        "p99_ms": 148.231,
        "min_ms": 140.334,
        "max_ms": 150.543,
        "implied_fps_at_mean": 6.94
      },
      "sr_output_shape": "1280x720",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 20.825,
        "p50_ms": 19.891,
        "p95_ms": 32.66,
        "p99_ms": 33.886,
        "min_ms": 9.6,
        "max_ms": 34.066,
        "implied_fps_at_mean": 48.02
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 77.145,
        "p50_ms": 73.437,
        "p95_ms": 96.25,
        "p99_ms": 105.176,
        "min_ms": 60.35,
        "max_ms": 120.149,
        "implied_fps_at_mean": 12.96
      },
      "attributed": {
        "per_frame_ms": 242.001,
        "implied_fps": 4.13,
        "model_share_pct": 51.3,
        "overhead_share_pct": 48.7
      }
    },
    {
      "input_w": 854,
      "input_h": 480,
      "model_forward": {
        "n": 40,
        "mean_ms": 220.425,
        "p50_ms": 220.114,
        "p95_ms": 222.864,
        "p99_ms": 222.871,
        "min_ms": 218.41,
        "max_ms": 226.818,
        "implied_fps_at_mean": 4.54
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 497.067,
        "p50_ms": 522.739,
        "p95_ms": 897.901,
        "p99_ms": 905.557,
        "min_ms": 265.173,
        "max_ms": 944.03,
        "implied_fps_at_mean": 2.01
      },
      "sr_output_shape": "1708x960",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 24.788,
        "p50_ms": 23.397,
        "p95_ms": 42.61,
        "p99_ms": 42.887,
        "min_ms": 14.58,
        "max_ms": 43.987,
        "implied_fps_at_mean": 40.34
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 60.727,
        "p50_ms": 59.916,
        "p95_ms": 76.117,
        "p99_ms": 78.977,
        "min_ms": 47.154,
        "max_ms": 79.962,
        "implied_fps_at_mean": 16.47
      },
      "attributed": {
        "per_frame_ms": 582.582,
        "implied_fps": 1.72,
        "model_share_pct": 37.8,
        "overhead_share_pct": 62.2
      }
    },
    {
      "input_w": 1280,
      "input_h": 720,
      "model_forward": {
        "n": 40,
        "mean_ms": 496.121,
        "p50_ms": 495.794,
        "p95_ms": 501.731,
        "p99_ms": 503.339,
        "min_ms": 491.433,
        "max_ms": 509.349,
        "implied_fps_at_mean": 2.02
      },
      "upscale_call": {
        "n": 40,
        "mean_ms": 862.912,
        "p50_ms": 600.043,
        "p95_ms": 1225.431,
        "p99_ms": 1232.333,
        "min_ms": 580.636,
        "max_ms": 1258.996,
        "implied_fps_at_mean": 1.16
      },
      "sr_output_shape": "2560x1440",
      "fit_to_canvas": {
        "n": 40,
        "mean_ms": 11.406,
        "p50_ms": 10.798,
        "p95_ms": 17.411,
        "p99_ms": 20.446,
        "min_ms": 6.194,
        "max_ms": 20.918,
        "implied_fps_at_mean": 87.68
      },
      "sharpness_probe": {
        "n": 40,
        "mean_ms": 58.702,
        "p50_ms": 57.239,
        "p95_ms": 71.944,
        "p99_ms": 75.153,
        "min_ms": 44.169,
        "max_ms": 80.843,
        "implied_fps_at_mean": 17.04
      },
      "attributed": {
        "per_frame_ms": 933.02,
        "implied_fps": 1.07,
        "model_share_pct": 53.2,
        "overhead_share_pct": 46.8
      }
    }
  ],
  "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": 1785922911.311304,
    "load_average": [
      26.27,
      44.44,
      57.28
    ],
    "cpu_count": 20,
    "mem_total_mib": 124611,
    "mem_available_mib": 55747,
    "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": 123.43,
  "arm": "quiet-uncapped"
}