{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 189838,
  "matchedSampleCount": 7997,
  "profileSpanCycles": 10798722,
  "profileSpanMs": 7.199148,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 28,
      "tileCount": 14,
      "phaseCount": 2,
      "phaseCycles": 61518,
      "phaseMs": 0.041012,
      "workCycles": 861168,
      "averageActiveTiles": 13.99863454598654,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 30756.0,
      "p50Cycles": 17136,
      "p95Cycles": 44376,
      "maximumCycles": 44382,
      "firstOffset": 343836,
      "lastOffset": 405354
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 2916,
      "tileCount": 1458,
      "phaseCount": 2,
      "phaseCycles": 50922,
      "phaseMs": 0.033948,
      "workCycles": 56086962,
      "averageActiveTiles": 1101.4288912454342,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 19234.211934156378,
      "p50Cycles": 16782,
      "p95Cycles": 30972,
      "maximumCycles": 32904,
      "firstOffset": 344940,
      "lastOffset": 419796
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 21,
      "phaseCycles": 29592,
      "phaseMs": 0.019728,
      "workCycles": 6508512,
      "averageActiveTiles": 219.94160583941607,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6508512,
      "usefulUtilization": 0.7082393026240099,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.10582288691420713,
      "meanCycles": 26784.0,
      "p50Cycles": 26784,
      "p95Cycles": 26784,
      "maximumCycles": 26784,
      "firstOffset": 375354,
      "lastOffset": 404946
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 34,
      "phaseCycles": 29406,
      "phaseMs": 0.019604,
      "workCycles": 32666490,
      "averageActiveTiles": 1110.8783921648644,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32666490,
      "usefulUtilization": 0.7233394221417728,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5458846020776594,
      "meanCycles": 26886.0,
      "p50Cycles": 26886,
      "p95Cycles": 26886,
      "maximumCycles": 26886,
      "firstOffset": 375684,
      "lastOffset": 405090
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 37,
      "phaseCycles": 13032,
      "phaseMs": 0.008688,
      "workCycles": 7262220,
      "averageActiveTiles": 557.2605893186004,
      "usefulCycles": 5458752.0,
      "estimatedWorkCycles": 7262220,
      "usefulUtilization": 0.7516643670943596,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.28456041316358394,
      "meanCycles": 4980.9465020576135,
      "p50Cycles": 5580,
      "p95Cycles": 6144,
      "maximumCycles": 6894,
      "firstOffset": 413100,
      "lastOffset": 426132
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 729,
      "tileCount": 729,
      "phaseCount": 29,
      "phaseCycles": 11466,
      "phaseMs": 0.007644,
      "workCycles": 6779700,
      "averageActiveTiles": 591.287284144427,
      "usefulCycles": 392202.0,
      "estimatedWorkCycles": 6779700,
      "usefulUtilization": 0.057849462365591395,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.023237534980547404,
      "meanCycles": 9300.0,
      "p50Cycles": 9300,
      "p95Cycles": 9300,
      "maximumCycles": 9300,
      "firstOffset": 346044,
      "lastOffset": 357510
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 153,
      "tileCount": 152,
      "phaseCount": 30,
      "phaseCycles": 8826,
      "phaseMs": 0.005884,
      "workCycles": 125010,
      "averageActiveTiles": 14.163834126444595,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 817.0588235294117,
      "p50Cycles": 858,
      "p95Cycles": 924,
      "maximumCycles": 1332,
      "firstOffset": 345078,
      "lastOffset": 420552
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 1251,
      "tileCount": 288,
      "phaseCount": 34,
      "phaseCycles": 6246,
      "phaseMs": 0.004164,
      "workCycles": 333666,
      "averageActiveTiles": 53.4207492795389,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 333666,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 266.71942446043164,
      "p50Cycles": 246,
      "p95Cycles": 246,
      "maximumCycles": 726,
      "firstOffset": 344922,
      "lastOffset": 358224
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 1488,
      "phaseMs": 0.000992,
      "workCycles": 1920,
      "averageActiveTiles": 1.2903225806451613,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 480.0,
      "p50Cycles": 432,
      "p95Cycles": 432,
      "maximumCycles": 624,
      "firstOffset": 416964,
      "lastOffset": 420168
    }
  ],
  "samples": []
}
