{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 40142,
  "matchedSampleCount": 40142,
  "profileSpanCycles": 805944,
  "profileSpanMs": 0.537296,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 1,
      "phaseCycles": 539478,
      "phaseMs": 0.359652,
      "workCycles": 787711692,
      "averageActiveTiles": 1460.1368211493332,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 535130.2255434783,
      "p50Cycles": 535128,
      "p95Cycles": 535140,
      "maximumCycles": 535302,
      "firstOffset": 266466,
      "lastOffset": 805944
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 88,
      "tileCount": 32,
      "phaseCount": 5,
      "phaseCycles": 266472,
      "phaseMs": 0.177648,
      "workCycles": 4843134,
      "averageActiveTiles": 18.1750202647933,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 55035.61363636364,
      "p50Cycles": 63684,
      "p95Cycles": 96732,
      "maximumCycles": 96738,
      "firstOffset": 0,
      "lastOffset": 266472
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 7272,
      "tileCount": 1458,
      "phaseCount": 5,
      "phaseCycles": 146526,
      "phaseMs": 0.097684,
      "workCycles": 154278318,
      "averageActiveTiles": 1052.9074566971049,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 21215.390264026402,
      "p50Cycles": 16140,
      "p95Cycles": 54804,
      "maximumCycles": 65976,
      "firstOffset": 0,
      "lastOffset": 268806
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k544_c32_r292_r292",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 20,
      "phaseCycles": 55680,
      "phaseMs": 0.03712,
      "workCycles": 65465280,
      "averageActiveTiles": 1175.7413793103449,
      "usefulCycles": 56477088.0,
      "estimatedWorkCycles": 65465280,
      "usefulUtilization": 0.8627029167216577,
      "workLaneOccupancy": 0.9876158340048348,
      "deviceUsefulUtilization": 0.6890730416041979,
      "meanCycles": 45462.0,
      "p50Cycles": 45462,
      "p95Cycles": 45462,
      "maximumCycles": 45462,
      "firstOffset": 197082,
      "lastOffset": 252762
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3523,
      "tileCount": 1458,
      "phaseCount": 23,
      "phaseCycles": 51648,
      "phaseMs": 0.034432,
      "workCycles": 27426660,
      "averageActiveTiles": 531.030436802974,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 7785.029804144196,
      "p50Cycles": 3132,
      "p95Cycles": 15402,
      "maximumCycles": 16590,
      "firstOffset": 22086,
      "lastOffset": 265956
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r160_r164",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 2,
      "phaseCycles": 50868,
      "phaseMs": 0.033912,
      "workCycles": 38516400,
      "averageActiveTiles": 757.1832979476292,
      "usefulCycles": 30700800.0,
      "estimatedWorkCycles": 38516400,
      "usefulUtilization": 0.7970838396111786,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.410012615772793,
      "meanCycles": 49380.0,
      "p50Cycles": 49380,
      "p95Cycles": 49380,
      "maximumCycles": 49380,
      "firstOffset": 38256,
      "lastOffset": 89124
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r160_r164",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 3,
      "phaseCycles": 50748,
      "phaseMs": 0.033832,
      "workCycles": 30232800,
      "averageActiveTiles": 595.7436746275715,
      "usefulCycles": 23886720.0,
      "estimatedWorkCycles": 30232800,
      "usefulUtilization": 0.7900928792569659,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.31976415433805916,
      "meanCycles": 48450.0,
      "p50Cycles": 48450,
      "p95Cycles": 48450,
      "maximumCycles": 48450,
      "firstOffset": 37770,
      "lastOffset": 88518
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r160_r164",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 2,
      "phaseCycles": 46710,
      "phaseMs": 0.03114,
      "workCycles": 1334160,
      "averageActiveTiles": 28.5626204238921,
      "usefulCycles": 1062720.0,
      "estimatedWorkCycles": 1334160,
      "usefulUtilization": 0.7965461413923367,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.015456144759989947,
      "meanCycles": 44472.0,
      "p50Cycles": 44472,
      "p95Cycles": 44472,
      "maximumCycles": 44472,
      "firstOffset": 37488,
      "lastOffset": 84198
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r160_r164",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 3,
      "phaseCycles": 45918,
      "phaseMs": 0.030612,
      "workCycles": 1047312,
      "averageActiveTiles": 22.808310466483732,
      "usefulCycles": 826848.0,
      "estimatedWorkCycles": 1047312,
      "usefulUtilization": 0.7894953939227279,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.01223305438617422,
      "meanCycles": 43638.0,
      "p50Cycles": 43638,
      "p95Cycles": 43638,
      "maximumCycles": 43638,
      "firstOffset": 37002,
      "lastOffset": 82920
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 6,
      "phaseCycles": 37470,
      "phaseMs": 0.02498,
      "workCycles": 34054008,
      "averageActiveTiles": 908.8339471577261,
      "usefulCycles": 18825696.0,
      "estimatedWorkCycles": 34054008,
      "usefulUtilization": 0.5528188047644789,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.34131827200891146,
      "meanCycles": 23356.65843621399,
      "p50Cycles": 24306,
      "p95Cycles": 24306,
      "maximumCycles": 24306,
      "firstOffset": 112332,
      "lastOffset": 149802
    }
  ],
  "samples": []
}
