{
  "generated_at": "2026-09-10T02:40:18.102316+00:00",
  "gpu_uuid": "GPU-46890c79-bf10-5b49-3d4a-a6b78f956122",
  "compute_references": {
    "bf16": {
      "three_launch_max_tflops": [
        787.0565134516788,
        787.7060668090745,
        789.0809256432037
      ],
      "reference_tflops": 787.7060668090745,
      "selection": "Median of three per-launch maxima across measured square GEMM sizes"
    },
    "fp32_ieee": {
      "three_launch_max_tflops": [
        51.286787913396516,
        51.275693333347874,
        51.250146619931485
      ],
      "reference_tflops": 51.275693333347874,
      "selection": "Median of three per-launch maxima across measured square GEMM sizes"
    },
    "fp32_tf32_allowed": {
      "three_launch_max_tflops": [
        397.8479222355699,
        397.44295597962594,
        397.4245633398784
      ],
      "reference_tflops": 397.44295597962594,
      "selection": "Median of three per-launch maxima across measured square GEMM sizes"
    }
  },
  "bandwidth_reference_GB_s": 4349.809747597248,
  "points": [
    {
      "input_sha256": "824a77bcaead8af0dacf908827027966a81c2d964d919eb4200bac0771b4a369",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 4680,
        "N": 5120,
        "K": 144
      },
      "arithmetic_intensity": 277.8535941783999,
      "nominal_tflops": 272.63514538558786,
      "reference_curve_tflops": 787.7060668090745,
      "ratio_to_empirical_reference": 0.3461127911455703,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:1"
    },
    {
      "input_sha256": "787b43c2c5ab957d3bd3c45047d646e6e407e506ea71b3249408d5d20bf00b0c",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 4680,
        "N": 5120,
        "K": 1536
      },
      "arithmetic_intensity": 921.6797607485263,
      "nominal_tflops": 770.6243216080403,
      "reference_curve_tflops": 787.7060668090745,
      "ratio_to_empirical_reference": 0.9783145694557986,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:2"
    },
    {
      "input_sha256": "ea1184eaad12555ac80bf5ee33de7aa2251926b289572cadadb4043884848587",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 3,
        "N": 5120,
        "K": 256
      },
      "arithmetic_intensity": 2.9684027442264953,
      "nominal_tflops": 1.683287671232877,
      "reference_curve_tflops": 12.911987191630832,
      "ratio_to_empirical_reference": 0.13036627486154373,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:3"
    },
    {
      "input_sha256": "a6cb8d5375b991c9c5d1a1e5674fa86eb61d10648b2ccf71bb7ec40c56f2ac1e",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 3,
        "N": 5120,
        "K": 5120
      },
      "arithmetic_intensity": 2.8240095972201154,
      "nominal_tflops": 9.068634686346863,
      "reference_curve_tflops": 12.283904473296236,
      "ratio_to_empirical_reference": 0.7382534361172385,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:4"
    },
    {
      "input_sha256": "d6a2e382ecf0fdad711df2cd1bcb239a004321ae666d26aa04920718cf49bf81",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 3,
        "N": 30720,
        "K": 5120
      },
      "arithmetic_intensity": 2.950628361218255,
      "nominal_tflops": 11.824859663191658,
      "reference_curve_tflops": 12.834672007164059,
      "ratio_to_empirical_reference": 0.9213215309741656,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:5"
    },
    {
      "input_sha256": "4f1e9988bf86b3f91f274f9b8c3bf33bf42b39945181ce6f09fbce8e0132cb6f",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 1170,
        "N": 15360,
        "K": 5120
      },
      "arithmetic_intensity": 619.8488421724303,
      "nominal_tflops": 727.8552082014935,
      "reference_curve_tflops": 787.7060668090745,
      "ratio_to_empirical_reference": 0.9240187918698768,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:6"
    },
    {
      "input_sha256": "a64f7da6eb1fcc9779d40258d974eb529c2397aaba4df6e32363d7c5d4e73065",
      "stage": "dit",
      "mode": "bf16",
      "shape_parameters": {
        "M": 1170,
        "N": 5120,
        "K": 5120
      },
      "arithmetic_intensity": 723.6115347331483,
      "nominal_tflops": 704.2351212343865,
      "reference_curve_tflops": 787.7060668090745,
      "ratio_to_empirical_reference": 0.894032877120242,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "bf16:7"
    },
    {
      "input_sha256": "2b9e80cee384305d79045df290612b04bcfe041bd67dc69a2bc096e45a6c5872",
      "stage": "vae",
      "mode": "fp32_tf32_allowed",
      "shape_parameters": {
        "input": [
          1,
          16,
          3,
          62,
          28
        ],
        "weight": [
          384,
          16,
          3,
          3,
          3
        ],
        "output": [
          1,
          384,
          1,
          60,
          26
        ],
        "stride": [
          1,
          1,
          1
        ],
        "padding": [
          0,
          0,
          0
        ],
        "dilation": [
          1,
          1,
          1
        ],
        "groups": 1
      },
      "arithmetic_intensity": 115.35117247675015,
      "nominal_tflops": 19.20912114014252,
      "reference_curve_tflops": 397.44295597962594,
      "ratio_to_empirical_reference": 0.0483317690026622,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "fp32_tf32_allowed:1"
    },
    {
      "input_sha256": "6309c12f24013edabd88edced7c2ce22672322d203206b12ba218c88696effb5",
      "stage": "vae",
      "mode": "fp32_tf32_allowed",
      "shape_parameters": {
        "input": [
          1,
          384,
          3,
          62,
          28
        ],
        "weight": [
          384,
          384,
          3,
          3,
          3
        ],
        "output": [
          1,
          384,
          1,
          60,
          26
        ],
        "stride": [
          1,
          1,
          1
        ],
        "padding": [
          0,
          0,
          0
        ],
        "dilation": [
          1,
          1,
          1
        ],
        "groups": 1
      },
      "arithmetic_intensity": 226.6271846655395,
      "nominal_tflops": 101.1143318572545,
      "reference_curve_tflops": 397.44295597962594,
      "ratio_to_empirical_reference": 0.254412187550351,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "fp32_tf32_allowed:2"
    },
    {
      "input_sha256": "3193a32ddc024e9d34eb5446f9909408dde38d963ebc5cb9654998ecf661a820",
      "stage": "vae",
      "mode": "fp32_tf32_allowed",
      "shape_parameters": {
        "input": [
          1,
          192,
          3,
          122,
          54
        ],
        "weight": [
          384,
          192,
          3,
          3,
          3
        ],
        "output": [
          1,
          384,
          1,
          120,
          52
        ],
        "stride": [
          1,
          1,
          1
        ],
        "padding": [
          0,
          0,
          0
        ],
        "dilation": [
          1,
          1,
          1
        ],
        "groups": 1
      },
      "arithmetic_intensity": 326.4852644327816,
      "nominal_tflops": 156.68129969727548,
      "reference_curve_tflops": 397.44295597962594,
      "ratio_to_empirical_reference": 0.3942233654917447,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "fp32_tf32_allowed:3"
    },
    {
      "input_sha256": "c7f10fe8b8ca84229e785645c70c003456c3950c79cdd176694e5bc86c9247f4",
      "stage": "vae",
      "mode": "fp32_tf32_allowed",
      "shape_parameters": {
        "input": [
          1,
          384,
          3,
          122,
          54
        ],
        "weight": [
          384,
          384,
          3,
          3,
          3
        ],
        "output": [
          1,
          384,
          1,
          120,
          52
        ],
        "stride": [
          1,
          1,
          1
        ],
        "padding": [
          0,
          0,
          0
        ],
        "dilation": [
          1,
          1,
          1
        ],
        "groups": 1
      },
      "arithmetic_intensity": 309.53992191686444,
      "nominal_tflops": 173.06193490860457,
      "reference_curve_tflops": 397.44295597962594,
      "ratio_to_empirical_reference": 0.4354384253258126,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "fp32_tf32_allowed:4"
    },
    {
      "input_sha256": "0cc0a21ce474602c53c3e5621e64598ed7be5a2c16603fc968441df41de71330",
      "stage": "vae",
      "mode": "fp32_tf32_allowed",
      "shape_parameters": {
        "input": [
          1,
          192,
          3,
          242,
          106
        ],
        "weight": [
          192,
          192,
          3,
          3,
          3
        ],
        "output": [
          1,
          192,
          1,
          240,
          104
        ],
        "stride": [
          1,
          1,
          1
        ],
        "padding": [
          0,
          0,
          0
        ],
        "dilation": [
          1,
          1,
          1
        ],
        "groups": 1
      },
      "arithmetic_intensity": 226.56157782675282,
      "nominal_tflops": 215.3552954230236,
      "reference_curve_tflops": 397.44295597962594,
      "ratio_to_empirical_reference": 0.5418520876592497,
      "scope": "Ratio to selected empirical reference, not MFU or hardware efficiency; no percentage-of-peak claim.",
      "plot_label": "fp32_tf32_allowed:5"
    }
  ],
  "excluded": [
    {
      "input_sha256": "4a5b492ee606c1a2c9fe4844a0c3125609fa599f4c027f56b4329c89d2879b7e",
      "reason": "different_physical_gpu"
    },
    {
      "input_sha256": "32bcd9b4d8be26c876626649d4ff2e623424efe309222bcd7e80afc47310ded4",
      "reason": "kernel_arithmetic_not_explicitly_identified_for_reference_selection"
    },
    {
      "input_sha256": "1e43418295ca6422f27afcf0b9940aa818fbb12fcf1013c12dfcf7a60f86e7d7",
      "reason": "kernel_arithmetic_not_explicitly_identified_for_reference_selection"
    },
    {
      "input_sha256": "1e47965836d719f164296b5c8d52cac8c22a10d9ff8c5f4fdfa48498909f8856",
      "reason": "different_physical_gpu"
    },
    {
      "input_sha256": "897dd2bd3ad5fe7258ee5c3ae9cfd93da4e5e94644902267f0b3d8bd156f63b0",
      "reason": "conversion_has_no_assigned_MAC_FLOPs"
    },
    {
      "input_sha256": "a9aab61b63e7cab483e81582c9af41d89d525814ab14de753e8a7b16de6c4f1b",
      "reason": "different_physical_gpu"
    }
  ],
  "source_sha256": {
    "same_gpu_gemm_calibration_audit.json": "c244cfdd2e60bd5ca5bf5cda74bc14f716dc359c96f13b2cc3e64d5021759b21",
    "local_bandwidth_audit.json": "2478941ab4958ba7f56312ec3df6e17a93a3c6d1bd6a3ef43e356b812602b87d",
    "local_bandwidth_gpu_identity.json": "02f5263025bb076acbcde2eda7080e78493c203e60b4bbb066f5c74d80983cce",
    "operator_ncu_physical_devices.json": "6f0f9e9ada2125d86c0dc6d5f577a3739cc22957ccb8cbfbd80dd3340bc7d535",
    "operator_work_traffic.json": "87c4c6d61626f3be61835643966ad4e0aefc75d04a7980a36748f114975758c6",
    "capability_operator_ncu_full_audit.json": "213f70d59c3357a7d3179a9d367cfe7a1ea050620462455b91c2f2bbb228dc13"
  },
  "scope": "Observed references, not theoretical ceilings. Operator points use summed NCU kernel time/DRAM traffic; references use unprofiled calibration. This diagnostic diagram mixes these only as explicitly labeled reference context, not an apples-to-apples speedup. Shapes/clock/kernel choices differ, so points can exceed a chosen empirical curve. No full-model MFU, L2 roofline, causal bottleneck proof or optimization acceptance."
}