{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 27117,
  "matchedSampleCount": 26009,
  "profileSpanCycles": 571104,
  "profileSpanMs": 0.380736,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 138,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 560742,
      "phaseMs": 0.373828,
      "workCycles": 15304644,
      "averageActiveTiles": 27.293557464930394,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 110903.21739130435,
      "p50Cycles": 60726,
      "p95Cycles": 245724,
      "maximumCycles": 246528,
      "firstOffset": 7746,
      "lastOffset": 568488
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 10166,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 198264,
      "phaseMs": 0.132176,
      "workCycles": 208311108,
      "averageActiveTiles": 1050.675402493645,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 20490.960849891795,
      "p50Cycles": 16992,
      "p95Cycles": 52698,
      "maximumCycles": 53292,
      "firstOffset": 0,
      "lastOffset": 570150
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k384_c432_r61",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 5,
      "phaseCycles": 132714,
      "phaseMs": 0.088476,
      "workCycles": 140583600,
      "averageActiveTiles": 1059.2974365929745,
      "usefulCycles": 85064256.0,
      "estimatedWorkCycles": 140583600,
      "usefulUtilization": 0.6050795114081585,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.4354342223978205,
      "meanCycles": 130170.0,
      "p50Cycles": 130170,
      "p95Cycles": 130170,
      "maximumCycles": 130170,
      "firstOffset": 121302,
      "lastOffset": 254016
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k384_c432_r60",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 5,
      "phaseCycles": 132324,
      "phaseMs": 0.088216,
      "workCycles": 46742400,
      "averageActiveTiles": 353.2420422599075,
      "usefulCycles": 27889920.0,
      "estimatedWorkCycles": 46742400,
      "usefulUtilization": 0.5966728280961183,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.14318609263428503,
      "meanCycles": 129840.0,
      "p50Cycles": 129840,
      "p95Cycles": 129840,
      "maximumCycles": 129840,
      "firstOffset": 114978,
      "lastOffset": 247302
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k288_c288_r122",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 10,
      "phaseCycles": 104808,
      "phaseMs": 0.069872,
      "workCycles": 75250080,
      "averageActiveTiles": 717.9803068468055,
      "usefulCycles": 56709504.0,
      "estimatedWorkCycles": 75250080,
      "usefulUtilization": 0.7536138699121648,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.3675814657361038,
      "meanCycles": 104514.0,
      "p50Cycles": 104514,
      "p95Cycles": 104514,
      "maximumCycles": 104514,
      "firstOffset": 425394,
      "lastOffset": 530202
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k288_c288_r121",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 10,
      "phaseCycles": 104580,
      "phaseMs": 0.06972,
      "workCycles": 75098880,
      "averageActiveTiles": 718.0998278829604,
      "usefulCycles": 56244672.0,
      "estimatedWorkCycles": 75098880,
      "usefulUtilization": 0.7489415554532903,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.36536331662051935,
      "meanCycles": 104304.0,
      "p50Cycles": 104304,
      "p95Cycles": 104304,
      "maximumCycles": 104304,
      "firstOffset": 425394,
      "lastOffset": 529974
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r61_p61_c384_p384",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 5,
      "phaseCycles": 73092,
      "phaseMs": 0.048728,
      "workCycles": 76191840,
      "averageActiveTiles": 1042.4101132818914,
      "usefulCycles": 6324480.0,
      "estimatedWorkCycles": 76191840,
      "usefulUtilization": 0.08300731416907638,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05878238027595954,
      "meanCycles": 70548.0,
      "p50Cycles": 70548,
      "p95Cycles": 70548,
      "maximumCycles": 70548,
      "firstOffset": 50754,
      "lastOffset": 123846
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r60_p60_c384_p384",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 5,
      "phaseCycles": 66654,
      "phaseMs": 0.044436,
      "workCycles": 23101200,
      "averageActiveTiles": 346.58385093167703,
      "usefulCycles": 2073600.0,
      "estimatedWorkCycles": 23101200,
      "usefulUtilization": 0.08976157082748948,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.02113445032817105,
      "meanCycles": 64170.0,
      "p50Cycles": 64170,
      "p95Cycles": 64170,
      "maximumCycles": 64170,
      "firstOffset": 50808,
      "lastOffset": 117462
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2880,
      "tileCount": 1440,
      "phaseCount": 16,
      "phaseCycles": 45132,
      "phaseMs": 0.030088,
      "workCycles": 50097816,
      "averageActiveTiles": 1110.0287157670832,
      "usefulCycles": 18032544.0,
      "estimatedWorkCycles": 50097816,
      "usefulUtilization": 0.3599467090541432,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.27143422771464576,
      "meanCycles": 17395.075,
      "p50Cycles": 16224,
      "p95Cycles": 18852,
      "maximumCycles": 18852,
      "firstOffset": 267300,
      "lastOffset": 568236
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 9,
      "phaseCycles": 44826,
      "phaseMs": 0.029884,
      "workCycles": 61454700,
      "averageActiveTiles": 1370.9610493909784,
      "usefulCycles": 37651392.0,
      "estimatedWorkCycles": 61454700,
      "usefulUtilization": 0.6126690391459074,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5706150739380678,
      "meanCycles": 42150.0,
      "p50Cycles": 42150,
      "p95Cycles": 42150,
      "maximumCycles": 42150,
      "firstOffset": 298068,
      "lastOffset": 342894
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c272_p288",
      "dimensions": {},
      "sampleCount": 48,
      "tileCount": 48,
      "phaseCount": 5,
      "phaseCycles": 34698,
      "phaseMs": 0.023132,
      "workCycles": 1365984,
      "averageActiveTiles": 39.367802178799934,
      "usefulCycles": 796416.0,
      "estimatedWorkCycles": 1365984,
      "usefulUtilization": 0.5830346475507766,
      "workLaneOccupancy": 0.9444444444444444,
      "deviceUsefulUtilization": 0.01559292980174274,
      "meanCycles": 28458.0,
      "p50Cycles": 28458,
      "p95Cycles": 28458,
      "maximumCycles": 28458,
      "firstOffset": 354138,
      "lastOffset": 388836
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c288_p288",
      "dimensions": {},
      "sampleCount": 672,
      "tileCount": 672,
      "phaseCount": 10,
      "phaseCycles": 33822,
      "phaseMs": 0.022548,
      "workCycles": 18277056,
      "averageActiveTiles": 540.3895689196381,
      "usefulCycles": 11805696.0,
      "estimatedWorkCycles": 18277056,
      "usefulUtilization": 0.6459298477829252,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.23712890760580327,
      "meanCycles": 27198.0,
      "p50Cycles": 27198,
      "p95Cycles": 27198,
      "maximumCycles": 27198,
      "firstOffset": 353790,
      "lastOffset": 387612
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c288_p288",
      "dimensions": {},
      "sampleCount": 672,
      "tileCount": 672,
      "phaseCount": 10,
      "phaseCycles": 33690,
      "phaseMs": 0.02246,
      "workCycles": 18277056,
      "averageActiveTiles": 542.506856634016,
      "usefulCycles": 11708928.0,
      "estimatedWorkCycles": 18277056,
      "usefulUtilization": 0.6406353408338848,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.23610670176932905,
      "meanCycles": 27198.0,
      "p50Cycles": 27198,
      "p95Cycles": 27198,
      "maximumCycles": 27198,
      "firstOffset": 353862,
      "lastOffset": 387552
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c272_p288",
      "dimensions": {},
      "sampleCount": 48,
      "tileCount": 48,
      "phaseCount": 5,
      "phaseCycles": 33216,
      "phaseMs": 0.022144,
      "workCycles": 1365984,
      "averageActiveTiles": 41.1242774566474,
      "usefulCycles": 789888.0,
      "estimatedWorkCycles": 1365984,
      "usefulUtilization": 0.5782556750298686,
      "workLaneOccupancy": 0.9444444444444444,
      "deviceUsefulUtilization": 0.01615512691631063,
      "meanCycles": 28458.0,
      "p50Cycles": 28458,
      "p95Cycles": 28458,
      "maximumCycles": 28458,
      "firstOffset": 354294,
      "lastOffset": 387510
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 10,
      "phaseCycles": 23322,
      "phaseMs": 0.015548,
      "workCycles": 22282560,
      "averageActiveTiles": 955.4309235914587,
      "usefulCycles": 6275232.0,
      "estimatedWorkCycles": 22282560,
      "usefulUtilization": 0.28162078324932144,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.18279157951253341,
      "meanCycles": 15474.0,
      "p50Cycles": 15474,
      "p95Cycles": 15474,
      "maximumCycles": 15474,
      "firstOffset": 380988,
      "lastOffset": 404310
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 2606,
      "tileCount": 1440,
      "phaseCount": 19,
      "phaseCycles": 18912,
      "phaseMs": 0.012608,
      "workCycles": 3954204,
      "averageActiveTiles": 209.08439086294416,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1517.3461243284728,
      "p50Cycles": 1554,
      "p95Cycles": 6126,
      "maximumCycles": 6222,
      "firstOffset": 264162,
      "lastOffset": 571104
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 1445,
      "tileCount": 1253,
      "phaseCount": 15,
      "phaseCycles": 13476,
      "phaseMs": 0.008984,
      "workCycles": 1117098,
      "averageActiveTiles": 82.89536954585931,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 773.0782006920415,
      "p50Cycles": 690,
      "p95Cycles": 690,
      "maximumCycles": 4638,
      "firstOffset": 0,
      "lastOffset": 343584
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 92,
      "tileCount": 83,
      "phaseCount": 3,
      "phaseCycles": 2592,
      "phaseMs": 0.001728,
      "workCycles": 50268,
      "averageActiveTiles": 19.39351851851852,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 546.3913043478261,
      "p50Cycles": 192,
      "p95Cycles": 1872,
      "maximumCycles": 2148,
      "firstOffset": 0,
      "lastOffset": 2592
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 1,
      "phaseCycles": 1002,
      "phaseMs": 0.000668,
      "workCycles": 12030,
      "averageActiveTiles": 12.005988023952096,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 1002,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 501.25,
      "p50Cycles": 558,
      "p95Cycles": 888,
      "maximumCycles": 1002,
      "firstOffset": 0,
      "lastOffset": 1002
    }
  ],
  "samples": []
}
