{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 21118,
  "matchedSampleCount": 21118,
  "profileSpanCycles": 128202,
  "profileSpanMs": 0.085468,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 223,
      "tileCount": 127,
      "phaseCount": 4,
      "phaseCycles": 100656,
      "phaseMs": 0.067104,
      "workCycles": 5619090,
      "averageActiveTiles": 55.82469003338102,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 25197.713004484303,
      "p50Cycles": 25248,
      "p95Cycles": 28434,
      "maximumCycles": 28440,
      "firstOffset": 0,
      "lastOffset": 100656
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 5665,
      "tileCount": 1440,
      "phaseCount": 4,
      "phaseCycles": 65544,
      "phaseMs": 0.043696,
      "workCycles": 52235184,
      "averageActiveTiles": 796.9483705602344,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 9220.68561341571,
      "p50Cycles": 8160,
      "p95Cycles": 16434,
      "maximumCycles": 31374,
      "firstOffset": 24912,
      "lastOffset": 119478
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k256_c16_r730_r730",
      "dimensions": {},
      "sampleCount": 1345,
      "tileCount": 1345,
      "phaseCount": 1,
      "phaseCycles": 24966,
      "phaseMs": 0.016644,
      "workCycles": 33544818,
      "averageActiveTiles": 1343.620043258832,
      "usefulCycles": 746496.0,
      "estimatedWorkCycles": 798912,
      "usefulUtilization": 0.9343907714491708,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.02031284285759067,
      "meanCycles": 24940.385130111525,
      "p50Cycles": 24942,
      "p95Cycles": 24960,
      "maximumCycles": 24966,
      "firstOffset": 0,
      "lastOffset": 24966
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c144_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 2,
      "phaseCycles": 24726,
      "phaseMs": 0.016484,
      "workCycles": 8881920,
      "averageActiveTiles": 359.2137830623635,
      "usefulCycles": 7127424.0,
      "estimatedWorkCycles": 8881920,
      "usefulUtilization": 0.8024643320363165,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.1958262557631643,
      "meanCycles": 24672.0,
      "p50Cycles": 24672,
      "p95Cycles": 24672,
      "maximumCycles": 24672,
      "firstOffset": 75678,
      "lastOffset": 100404
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c144_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 5,
      "phaseCycles": 24552,
      "phaseMs": 0.016368,
      "workCycles": 26412480,
      "averageActiveTiles": 1075.7771260997067,
      "usefulCycles": 21111120.0,
      "estimatedWorkCycles": 26412480,
      "usefulUtilization": 0.7992857921709737,
      "workLaneOccupancy": 0.9944715778049111,
      "deviceUsefulUtilization": 0.584139519316588,
      "meanCycles": 24456.0,
      "p50Cycles": 24456,
      "p95Cycles": 24456,
      "maximumCycles": 24456,
      "firstOffset": 75738,
      "lastOffset": 100290
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 4699,
      "tileCount": 1440,
      "phaseCount": 14,
      "phaseCycles": 20166,
      "phaseMs": 0.013444,
      "workCycles": 4870530,
      "averageActiveTiles": 241.5218684915204,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1036.5035113854012,
      "p50Cycles": 1422,
      "p95Cycles": 1422,
      "maximumCycles": 1422,
      "firstOffset": 29994,
      "lastOffset": 128202
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2785,
      "tileCount": 1440,
      "phaseCount": 7,
      "phaseCycles": 19440,
      "phaseMs": 0.01296,
      "workCycles": 18902412,
      "averageActiveTiles": 972.3462962962963,
      "usefulCycles": 12392480.0,
      "estimatedWorkCycles": 18902412,
      "usefulUtilization": 0.6556031050428909,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.43306606727500446,
      "meanCycles": 6787.221543985637,
      "p50Cycles": 7410,
      "p95Cycles": 8190,
      "maximumCycles": 8190,
      "firstOffset": 31416,
      "lastOffset": 127422
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 5,
      "phaseCycles": 14712,
      "phaseMs": 0.009808,
      "workCycles": 12925440,
      "averageActiveTiles": 878.5644371941272,
      "usefulCycles": 3137616.0,
      "estimatedWorkCycles": 12925440,
      "usefulUtilization": 0.24274732620320855,
      "workLaneOccupancy": 0.9949315068493151,
      "deviceUsefulUtilization": 0.14488394566990567,
      "meanCycles": 8976.0,
      "p50Cycles": 8976,
      "p95Cycles": 8976,
      "maximumCycles": 8976,
      "firstOffset": 57282,
      "lastOffset": 71994
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1345,
      "tileCount": 1345,
      "phaseCount": 1,
      "phaseCycles": 13182,
      "phaseMs": 0.008788,
      "workCycles": 17592600,
      "averageActiveTiles": 1334.5926263086026,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 17592600,
      "usefulUtilization": 0.535045871559633,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.4851007302448002,
      "meanCycles": 13080.0,
      "p50Cycles": 13080,
      "p95Cycles": 13080,
      "maximumCycles": 13080,
      "firstOffset": 38706,
      "lastOffset": 51888
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 2176,
      "tileCount": 384,
      "phaseCount": 11,
      "phaseCycles": 3456,
      "phaseMs": 0.002304,
      "workCycles": 545520,
      "averageActiveTiles": 157.84722222222223,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 545520,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 250.6985294117647,
      "p50Cycles": 240,
      "p95Cycles": 240,
      "maximumCycles": 972,
      "firstOffset": 25236,
      "lastOffset": 53298
    }
  ],
  "samples": []
}
