{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 5536,
  "matchedSampleCount": 5457,
  "profileSpanCycles": 67212,
  "profileSpanMs": 0.044808,
  "atOffset": null,
  "groups": [
    {
      "name": "idle",
      "dimensions": {},
      "sampleCount": 1408,
      "tileCount": 1408,
      "phaseCount": 1,
      "phaseCycles": 65652,
      "phaseMs": 0.043768,
      "workCycles": 92400360,
      "averageActiveTiles": 1407.4264302686895,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 65625.25568181818,
      "p50Cycles": 65628,
      "p95Cycles": 65646,
      "maximumCycles": 65652,
      "firstOffset": 0,
      "lastOffset": 65652
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 64,
      "phaseCount": 18,
      "phaseCycles": 27306,
      "phaseMs": 0.018204,
      "workCycles": 1452792,
      "averageActiveTiles": 53.20413096022852,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1261.1041666666667,
      "p50Cycles": 1308,
      "p95Cycles": 1782,
      "maximumCycles": 2070,
      "firstOffset": 0,
      "lastOffset": 66012
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 1104,
      "tileCount": 64,
      "phaseCount": 60,
      "phaseCycles": 15342,
      "phaseMs": 0.010228,
      "workCycles": 507870,
      "averageActiveTiles": 33.103245991396165,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 460.0271739130435,
      "p50Cycles": 504,
      "p95Cycles": 528,
      "maximumCycles": 528,
      "firstOffset": 1680,
      "lastOffset": 66408
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 192,
      "tileCount": 64,
      "phaseCount": 9,
      "phaseCycles": 11466,
      "phaseMs": 0.007644,
      "workCycles": 652032,
      "averageActiveTiles": 56.866562009419155,
      "usefulCycles": 297216.0,
      "estimatedWorkCycles": 652032,
      "usefulUtilization": 0.4558303886925795,
      "workLaneOccupancy": 0.9923076923076923,
      "deviceUsefulUtilization": 0.01760971947307351,
      "meanCycles": 3396.0,
      "p50Cycles": 3396,
      "p95Cycles": 3396,
      "maximumCycles": 3396,
      "firstOffset": 7422,
      "lastOffset": 56124
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c64_r32_r34",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 48,
      "phaseCount": 18,
      "phaseCycles": 10656,
      "phaseMs": 0.007104,
      "workCycles": 452736,
      "averageActiveTiles": 42.486486486486484,
      "usefulCycles": 145920.0,
      "estimatedWorkCycles": 452736,
      "usefulUtilization": 0.32230703986429177,
      "workLaneOccupancy": 0.9895833333333334,
      "deviceUsefulUtilization": 0.009302781041911477,
      "meanCycles": 1572.0,
      "p50Cycles": 1572,
      "p95Cycles": 1572,
      "maximumCycles": 1572,
      "firstOffset": 2250,
      "lastOffset": 61458
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 64,
      "phaseCount": 27,
      "phaseCycles": 10494,
      "phaseMs": 0.006996,
      "workCycles": 503712,
      "averageActiveTiles": 48.0,
      "usefulCycles": 162240.0,
      "estimatedWorkCycles": 503712,
      "usefulUtilization": 0.3220888126548504,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.010502896064832078,
      "meanCycles": 1311.75,
      "p50Cycles": 1188,
      "p95Cycles": 1512,
      "maximumCycles": 1512,
      "firstOffset": 5496,
      "lastOffset": 64488
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k32_c64_r32_r34",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 16,
      "phaseCount": 9,
      "phaseCycles": 9942,
      "phaseMs": 0.006628,
      "workCycles": 153792,
      "averageActiveTiles": 15.468919734459867,
      "usefulCycles": 52224.0,
      "estimatedWorkCycles": 153792,
      "usefulUtilization": 0.33957553058676654,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0035685235233921964,
      "meanCycles": 1602.0,
      "p50Cycles": 1602,
      "p95Cycles": 1602,
      "maximumCycles": 1602,
      "firstOffset": 2208,
      "lastOffset": 61470
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 64,
      "phaseCount": 30,
      "phaseCycles": 7086,
      "phaseMs": 0.004724,
      "workCycles": 387984,
      "averageActiveTiles": 54.75359864521592,
      "usefulCycles": 30960.0,
      "estimatedWorkCycles": 387984,
      "usefulUtilization": 0.07979710503525919,
      "workLaneOccupancy": 0.9938366718027735,
      "deviceUsefulUtilization": 0.0029681920259176085,
      "meanCycles": 1010.375,
      "p50Cycles": 810,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 12804,
      "lastOffset": 67212
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 294,
      "tileCount": 61,
      "phaseCount": 23,
      "phaseCycles": 5328,
      "phaseMs": 0.003552,
      "workCycles": 141120,
      "averageActiveTiles": 26.486486486486488,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 480.0,
      "p50Cycles": 480,
      "p95Cycles": 480,
      "maximumCycles": 480,
      "firstOffset": 0,
      "lastOffset": 65370
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 16,
      "phaseCount": 11,
      "phaseCycles": 3168,
      "phaseMs": 0.002112,
      "workCycles": 31488,
      "averageActiveTiles": 9.93939393939394,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 31488,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 246.0,
      "p50Cycles": 246,
      "p95Cycles": 246,
      "maximumCycles": 246,
      "firstOffset": 10836,
      "lastOffset": 56484
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 27,
      "tileCount": 7,
      "phaseCount": 16,
      "phaseCycles": 2640,
      "phaseMs": 0.00176,
      "workCycles": 7938,
      "averageActiveTiles": 3.006818181818182,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 294.0,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 294,
      "firstOffset": 19758,
      "lastOffset": 65142
    }
  ],
  "samples": []
}
