{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 203140,
  "matchedSampleCount": 203140,
  "profileSpanCycles": 973806,
  "profileSpanMs": 0.649204,
  "atOffset": null,
  "groups": [
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 63103,
      "tileCount": 1472,
      "phaseCount": 49,
      "phaseCycles": 576354,
      "phaseMs": 0.384236,
      "workCycles": 539537820,
      "averageActiveTiles": 936.1222790160214,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 8550.113623757983,
      "p50Cycles": 7344,
      "p95Cycles": 20994,
      "maximumCycles": 45444,
      "firstOffset": 3732,
      "lastOffset": 972582
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 9025,
      "tileCount": 1440,
      "phaseCount": 31,
      "phaseCycles": 373662,
      "phaseMs": 0.249108,
      "workCycles": 30934602,
      "averageActiveTiles": 82.78765836504648,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 3427.656731301939,
      "p50Cycles": 1902,
      "p95Cycles": 7074,
      "maximumCycles": 51510,
      "firstOffset": 3738,
      "lastOffset": 972468
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 13821,
      "tileCount": 1472,
      "phaseCount": 147,
      "phaseCycles": 162270,
      "phaseMs": 0.10818,
      "workCycles": 191334468,
      "averageActiveTiles": 1179.1117766685154,
      "usefulCycles": 38793648.0,
      "estimatedWorkCycles": 185918400,
      "usefulUtilization": 0.208659540959905,
      "workLaneOccupancy": 0.9900654652722085,
      "deviceUsefulUtilization": 0.16241067624812108,
      "meanCycles": 13843.749945734751,
      "p50Cycles": 15264,
      "p95Cycles": 33318,
      "maximumCycles": 33750,
      "firstOffset": 0,
      "lastOffset": 968082
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 25874,
      "tileCount": 1472,
      "phaseCount": 177,
      "phaseCycles": 107298,
      "phaseMs": 0.071532,
      "workCycles": 34326288,
      "averageActiveTiles": 319.91545042778057,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1326.6710983999383,
      "p50Cycles": 1194,
      "p95Cycles": 2406,
      "maximumCycles": 3186,
      "firstOffset": 15942,
      "lastOffset": 972084
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10637,
      "tileCount": 1472,
      "phaseCount": 143,
      "phaseCycles": 72558,
      "phaseMs": 0.048372,
      "workCycles": 58171176,
      "averageActiveTiles": 801.7196725378318,
      "usefulCycles": 24901424.0,
      "estimatedWorkCycles": 58171176,
      "usefulUtilization": 0.4280715246327494,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.23314766477672436,
      "meanCycles": 5468.757732443358,
      "p50Cycles": 5178,
      "p95Cycles": 9666,
      "maximumCycles": 10428,
      "firstOffset": 16920,
      "lastOffset": 971802
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 15936,
      "tileCount": 1472,
      "phaseCount": 105,
      "phaseCycles": 61566,
      "phaseMs": 0.041044,
      "workCycles": 6780096,
      "averageActiveTiles": 110.12727804307572,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 2803200,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 425.4578313253012,
      "p50Cycles": 300,
      "p95Cycles": 960,
      "maximumCycles": 1500,
      "firstOffset": 171702,
      "lastOffset": 929466
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 6784,
      "tileCount": 1472,
      "phaseCount": 97,
      "phaseCycles": 60378,
      "phaseMs": 0.040252,
      "workCycles": 8984088,
      "averageActiveTiles": 148.7973765278744,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1324.305424528302,
      "p50Cycles": 978,
      "p95Cycles": 3054,
      "maximumCycles": 3054,
      "firstOffset": 145950,
      "lastOffset": 933540
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1504,
      "tileCount": 1472,
      "phaseCount": 26,
      "phaseCycles": 56604,
      "phaseMs": 0.037736,
      "workCycles": 69399510,
      "averageActiveTiles": 1226.053105787577,
      "usefulCycles": 31442880.0,
      "estimatedWorkCycles": 69399510,
      "usefulUtilization": 0.4530706340722002,
      "workLaneOccupancy": 0.9970370671023125,
      "deviceUsefulUtilization": 0.3773700122590814,
      "meanCycles": 46143.29122340425,
      "p50Cycles": 47010,
      "p95Cycles": 47010,
      "maximumCycles": 47010,
      "firstOffset": 251160,
      "lastOffset": 917070
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r180_r184",
      "dimensions": {},
      "sampleCount": 920,
      "tileCount": 920,
      "phaseCount": 22,
      "phaseCycles": 48906,
      "phaseMs": 0.032604,
      "workCycles": 44905200,
      "averageActiveTiles": 918.1940866151392,
      "usefulCycles": 35637120.0,
      "estimatedWorkCycles": 44905200,
      "usefulUtilization": 0.7936078672403196,
      "workLaneOccupancy": 0.9746376811594203,
      "deviceUsefulUtilization": 0.4950312845049687,
      "meanCycles": 48810.0,
      "p50Cycles": 48810,
      "p95Cycles": 48810,
      "maximumCycles": 48810,
      "firstOffset": 649716,
      "lastOffset": 698622
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c96_r180_r184",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 20,
      "phaseCycles": 48834,
      "phaseMs": 0.032556,
      "workCycles": 38015640,
      "averageActiveTiles": 778.4666420936232,
      "usefulCycles": 31000320.0,
      "estimatedWorkCycles": 38015640,
      "usefulUtilization": 0.8154622676351102,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4312569111684482,
      "meanCycles": 48738.0,
      "p50Cycles": 48738,
      "p95Cycles": 48738,
      "maximumCycles": 48738,
      "firstOffset": 488898,
      "lastOffset": 537732
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r180_r184",
      "dimensions": {},
      "sampleCount": 552,
      "tileCount": 552,
      "phaseCount": 28,
      "phaseCycles": 48276,
      "phaseMs": 0.032184,
      "workCycles": 26598672,
      "averageActiveTiles": 550.9709172259508,
      "usefulCycles": 20839968.0,
      "estimatedWorkCycles": 26598672,
      "usefulUtilization": 0.7834965595274832,
      "workLaneOccupancy": 0.971027911969941,
      "deviceUsefulUtilization": 0.29326346334662645,
      "meanCycles": 48186.0,
      "p50Cycles": 48186,
      "p95Cycles": 48186,
      "maximumCycles": 48186,
      "firstOffset": 649722,
      "lastOffset": 697998
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c96_r180_r184",
      "dimensions": {},
      "sampleCount": 468,
      "tileCount": 468,
      "phaseCount": 21,
      "phaseCycles": 48192,
      "phaseMs": 0.032128,
      "workCycles": 22517352,
      "averageActiveTiles": 467.24252988047806,
      "usefulCycles": 18128448.0,
      "estimatedWorkCycles": 22517352,
      "usefulUtilization": 0.8050879161990273,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.255551164905595,
      "meanCycles": 48114.0,
      "p50Cycles": 48114,
      "p95Cycles": 48114,
      "maximumCycles": 48114,
      "firstOffset": 488910,
      "lastOffset": 537102
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c80_r180_r184",
      "dimensions": {},
      "sampleCount": 140,
      "tileCount": 140,
      "phaseCount": 12,
      "phaseCycles": 41814,
      "phaseMs": 0.027876,
      "workCycles": 5693520,
      "averageActiveTiles": 136.16300760510833,
      "usefulCycles": 4636800.0,
      "estimatedWorkCycles": 5693520,
      "usefulUtilization": 0.8143995278843317,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07533362031855359,
      "meanCycles": 40668.0,
      "p50Cycles": 40668,
      "p95Cycles": 40668,
      "maximumCycles": 40668,
      "firstOffset": 487758,
      "lastOffset": 529572
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 12318,
      "tileCount": 1458,
      "phaseCount": 148,
      "phaseCycles": 41472,
      "phaseMs": 0.027648,
      "workCycles": 27526248,
      "averageActiveTiles": 663.7309027777778,
      "usefulCycles": 5350248.0,
      "estimatedWorkCycles": 27526248,
      "usefulUtilization": 0.1943689528627367,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.08764176668176328,
      "meanCycles": 2234.6361422308814,
      "p50Cycles": 1644,
      "p95Cycles": 4812,
      "maximumCycles": 4812,
      "firstOffset": 30426,
      "lastOffset": 973806
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c80_r180_r184",
      "dimensions": {},
      "sampleCount": 84,
      "tileCount": 84,
      "phaseCount": 16,
      "phaseCycles": 40212,
      "phaseMs": 0.026808,
      "workCycles": 3372264,
      "averageActiveTiles": 83.86213070725157,
      "usefulCycles": 2711520.0,
      "estimatedWorkCycles": 3372264,
      "usefulUtilization": 0.8040651621581228,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.04580884356389397,
      "meanCycles": 40146.0,
      "p50Cycles": 40146,
      "p95Cycles": 40146,
      "maximumCycles": 40146,
      "firstOffset": 487764,
      "lastOffset": 527976
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 2,
      "phaseCycles": 39594,
      "phaseMs": 0.026396,
      "workCycles": 31954500,
      "averageActiveTiles": 807.0540991059252,
      "usefulCycles": 25505280.0,
      "estimatedWorkCycles": 31954500,
      "usefulUtilization": 0.7981749049429658,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4376157125256132,
      "meanCycles": 39450.0,
      "p50Cycles": 39450,
      "p95Cycles": 39450,
      "maximumCycles": 39450,
      "firstOffset": 102570,
      "lastOffset": 142164
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 2,
      "phaseCycles": 38784,
      "phaseMs": 0.025856,
      "workCycles": 25089264,
      "averageActiveTiles": 646.8972772277227,
      "usefulCycles": 19844352.0,
      "estimatedWorkCycles": 25089264,
      "usefulUtilization": 0.7909499457616612,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.3475973956091261,
      "meanCycles": 38718.0,
      "p50Cycles": 38718,
      "p95Cycles": 38718,
      "maximumCycles": 38718,
      "firstOffset": 102642,
      "lastOffset": 141426
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c64_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 22,
      "phaseCycles": 31734,
      "phaseMs": 0.021156,
      "workCycles": 19738080,
      "averageActiveTiles": 621.9852524106636,
      "usefulCycles": 15344640.0,
      "estimatedWorkCycles": 19738080,
      "usefulUtilization": 0.7774130006565988,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.32849145478310193,
      "meanCycles": 27414.0,
      "p50Cycles": 27414,
      "p95Cycles": 27414,
      "maximumCycles": 27414,
      "firstOffset": 803892,
      "lastOffset": 835626
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c64_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 26,
      "phaseCycles": 31296,
      "phaseMs": 0.020864,
      "workCycles": 19431360,
      "averageActiveTiles": 620.8895705521472,
      "usefulCycles": 14888448.0,
      "estimatedWorkCycles": 19431360,
      "usefulUtilization": 0.7662072032014229,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.3231861829821286,
      "meanCycles": 26988.0,
      "p50Cycles": 26988,
      "p95Cycles": 26988,
      "maximumCycles": 26988,
      "firstOffset": 803898,
      "lastOffset": 835194
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r182_r183",
      "dimensions": {},
      "sampleCount": 768,
      "tileCount": 768,
      "phaseCount": 9,
      "phaseCycles": 30378,
      "phaseMs": 0.020252,
      "workCycles": 20731392,
      "averageActiveTiles": 682.4475607347423,
      "usefulCycles": 14329224.0,
      "estimatedWorkCycles": 20731392,
      "usefulUtilization": 0.6911848466325851,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3204466118214841,
      "meanCycles": 26994.0,
      "p50Cycles": 26994,
      "p95Cycles": 26994,
      "maximumCycles": 26994,
      "firstOffset": 314178,
      "lastOffset": 344556
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 35,
      "phaseCycles": 30246,
      "phaseMs": 0.020164,
      "workCycles": 33942240,
      "averageActiveTiles": 1122.205911525491,
      "usefulCycles": 18825696.0,
      "estimatedWorkCycles": 33942240,
      "usefulUtilization": 0.554639175257732,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.42283923997136524,
      "meanCycles": 23280.0,
      "p50Cycles": 23280,
      "p95Cycles": 23280,
      "maximumCycles": 23280,
      "firstOffset": 574878,
      "lastOffset": 605124
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r182_r183",
      "dimensions": {},
      "sampleCount": 256,
      "tileCount": 256,
      "phaseCount": 9,
      "phaseCycles": 29544,
      "phaseMs": 0.019696,
      "workCycles": 6931968,
      "averageActiveTiles": 234.63200649878146,
      "usefulCycles": 4802652.0,
      "estimatedWorkCycles": 6931968,
      "usefulUtilization": 0.6928266258586306,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.11043430800339067,
      "meanCycles": 27078.0,
      "p50Cycles": 27078,
      "p95Cycles": 27078,
      "maximumCycles": 27078,
      "firstOffset": 315264,
      "lastOffset": 344808
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r182_r183",
      "dimensions": {},
      "sampleCount": 192,
      "tileCount": 192,
      "phaseCount": 6,
      "phaseCycles": 28800,
      "phaseMs": 0.0192,
      "workCycles": 4333824,
      "averageActiveTiles": 150.48,
      "usefulCycles": 3144960.0,
      "estimatedWorkCycles": 4333824,
      "usefulUtilization": 0.7256778309409888,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.07418478260869565,
      "meanCycles": 22572.0,
      "p50Cycles": 22572,
      "p95Cycles": 22572,
      "maximumCycles": 22572,
      "firstOffset": 214758,
      "lastOffset": 243558
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r128_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 1600,
      "tileCount": 800,
      "phaseCount": 39,
      "phaseCycles": 25236,
      "phaseMs": 0.016824,
      "workCycles": 6816000,
      "averageActiveTiles": 270.0903471231574,
      "usefulCycles": 1638400.0,
      "estimatedWorkCycles": 6816000,
      "usefulUtilization": 0.2403755868544601,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.04410538430261807,
      "meanCycles": 4260.0,
      "p50Cycles": 4260,
      "p95Cycles": 4260,
      "maximumCycles": 4260,
      "firstOffset": 189042,
      "lastOffset": 897744
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c80_r182_r183",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 7,
      "phaseCycles": 24840,
      "phaseMs": 0.01656,
      "workCycles": 1448832,
      "averageActiveTiles": 58.32657004830918,
      "usefulCycles": 1054080.0,
      "estimatedWorkCycles": 1448832,
      "usefulUtilization": 0.7275377683540949,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.028827977315689982,
      "meanCycles": 22638.0,
      "p50Cycles": 22638,
      "p95Cycles": 22638,
      "maximumCycles": 22638,
      "firstOffset": 218820,
      "lastOffset": 243660
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r182_r183",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 8,
      "phaseCycles": 21294,
      "phaseMs": 0.014196,
      "workCycles": 15655680,
      "averageActiveTiles": 735.2155536770921,
      "usefulCycles": 11184264.0,
      "estimatedWorkCycles": 15655680,
      "usefulUtilization": 0.7143901766004415,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.35681438127090304,
      "meanCycles": 18120.0,
      "p50Cycles": 18120,
      "p95Cycles": 18120,
      "maximumCycles": 18120,
      "firstOffset": 211914,
      "lastOffset": 233208
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r182_r183",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 288,
      "phaseCount": 6,
      "phaseCycles": 20688,
      "phaseMs": 0.013792,
      "workCycles": 5234112,
      "averageActiveTiles": 253.00232018561485,
      "usefulCycles": 3748572.0,
      "estimatedWorkCycles": 5234112,
      "usefulUtilization": 0.7161810828656322,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.12309475247150206,
      "meanCycles": 18174.0,
      "p50Cycles": 18174,
      "p95Cycles": 18174,
      "maximumCycles": 18174,
      "firstOffset": 218406,
      "lastOffset": 239094
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 2980,
      "tileCount": 1458,
      "phaseCount": 74,
      "phaseCycles": 19800,
      "phaseMs": 0.0132,
      "workCycles": 19716000,
      "averageActiveTiles": 995.7575757575758,
      "usefulCycles": 6723072.0,
      "estimatedWorkCycles": 19716000,
      "usefulUtilization": 0.34099573950091294,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.23067193675889328,
      "meanCycles": 6616.107382550335,
      "p50Cycles": 6696,
      "p95Cycles": 6696,
      "maximumCycles": 6696,
      "firstOffset": 430380,
      "lastOffset": 948024
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r89_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 320,
      "tileCount": 160,
      "phaseCount": 24,
      "phaseCycles": 19122,
      "phaseMs": 0.012748,
      "workCycles": 1213440,
      "averageActiveTiles": 63.4577973015375,
      "usefulCycles": 227840.0,
      "estimatedWorkCycles": 1213440,
      "usefulUtilization": 0.1877637130801688,
      "workLaneOccupancy": 0.6953125,
      "deviceUsefulUtilization": 0.008094478019854209,
      "meanCycles": 3792.0,
      "p50Cycles": 3792,
      "p95Cycles": 3792,
      "maximumCycles": 3792,
      "firstOffset": 189888,
      "lastOffset": 896928
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c64_r72_r76",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 18,
      "phaseCycles": 17082,
      "phaseMs": 0.011388,
      "workCycles": 6136560,
      "averageActiveTiles": 359.2413066385669,
      "usefulCycles": 3939840.0,
      "estimatedWorkCycles": 6136560,
      "usefulUtilization": 0.6420274551214361,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.15668667246987675,
      "meanCycles": 17046.0,
      "p50Cycles": 17046,
      "p95Cycles": 17046,
      "maximumCycles": 17046,
      "firstOffset": 398550,
      "lastOffset": 415632
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c64_r72_r76",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 25,
      "phaseCycles": 16686,
      "phaseMs": 0.011124,
      "workCycles": 17949600,
      "averageActiveTiles": 1075.7281553398059,
      "usefulCycles": 11176704.0,
      "estimatedWorkCycles": 17949600,
      "usefulUtilization": 0.6226714801444043,
      "workLaneOccupancy": 0.9981481481481481,
      "deviceUsefulUtilization": 0.4550443224989447,
      "meanCycles": 16620.0,
      "p50Cycles": 16620,
      "p95Cycles": 16620,
      "maximumCycles": 16620,
      "firstOffset": 398556,
      "lastOffset": 415242
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 352,
      "tileCount": 352,
      "phaseCount": 24,
      "phaseCycles": 14844,
      "phaseMs": 0.009896,
      "workCycles": 4145856,
      "averageActiveTiles": 279.2950687146322,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4145856,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.041240495354586244,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 878640,
      "lastOffset": 893484
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 3094,
      "tileCount": 1041,
      "phaseCount": 113,
      "phaseCycles": 12510,
      "phaseMs": 0.00834,
      "workCycles": 602952,
      "averageActiveTiles": 48.197601918465224,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 194.87782805429865,
      "p50Cycles": 198,
      "p95Cycles": 288,
      "maximumCycles": 1068,
      "firstOffset": 185538,
      "lastOffset": 972774
    },
    {
      "name": "add_layer_norm_moments",
      "dimensions": {},
      "sampleCount": 2916,
      "tileCount": 1458,
      "phaseCount": 58,
      "phaseCycles": 11484,
      "phaseMs": 0.007656,
      "workCycles": 11809800,
      "averageActiveTiles": 1028.369905956113,
      "usefulCycles": 3779136.0,
      "estimatedWorkCycles": 11809800,
      "usefulUtilization": 0.32,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2235586752078506,
      "meanCycles": 4050.0,
      "p50Cycles": 4050,
      "p95Cycles": 4050,
      "maximumCycles": 4050,
      "firstOffset": 426330,
      "lastOffset": 754752
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 2,
      "phaseCycles": 10962,
      "phaseMs": 0.007308,
      "workCycles": 13673124,
      "averageActiveTiles": 1247.320197044335,
      "usefulCycles": 4828896.0,
      "estimatedWorkCycles": 13673124,
      "usefulUtilization": 0.3531669865642994,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.29926108374384236,
      "meanCycles": 9378.0,
      "p50Cycles": 9378,
      "p95Cycles": 9378,
      "maximumCycles": 9378,
      "firstOffset": 33714,
      "lastOffset": 44676
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1504,
      "tileCount": 1472,
      "phaseCount": 37,
      "phaseCycles": 10374,
      "phaseMs": 0.006916,
      "workCycles": 6454080,
      "averageActiveTiles": 622.13996529786,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 4291.276595744681,
      "p50Cycles": 4350,
      "p95Cycles": 4350,
      "maximumCycles": 4350,
      "firstOffset": 358116,
      "lastOffset": 931338
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 48,
      "phaseCycles": 9750,
      "phaseMs": 0.0065,
      "workCycles": 1713408,
      "averageActiveTiles": 175.73415384615384,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1713408,
      "usefulUtilization": 0.015316842223218288,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0018285953177257525,
      "meanCycles": 1164.0,
      "p50Cycles": 1164,
      "p95Cycles": 1164,
      "maximumCycles": 1164,
      "firstOffset": 899802,
      "lastOffset": 909588
    },
    {
      "name": "gemm_f8_init_small_rows_k160_c16_r144_r144",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 1,
      "phaseCycles": 9720,
      "phaseMs": 0.00648,
      "workCycles": 12135168,
      "averageActiveTiles": 1248.474074074074,
      "usefulCycles": 15431472.0,
      "estimatedWorkCycles": 12135168,
      "usefulUtilization": 1.2716323333966204,
      "workLaneOccupancy": 0.9100118885869565,
      "deviceUsefulUtilization": 1.0785326086956522,
      "meanCycles": 8244.0,
      "p50Cycles": 8244,
      "p95Cycles": 8244,
      "maximumCycles": 8244,
      "firstOffset": 4146,
      "lastOffset": 13866
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 8,
      "phaseCycles": 9624,
      "phaseMs": 0.006416,
      "workCycles": 293184,
      "averageActiveTiles": 30.463840399002493,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 293184,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0022588456395243777,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 881256,
      "lastOffset": 890880
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c32_p32",
      "dimensions": {},
      "sampleCount": 704,
      "tileCount": 704,
      "phaseCount": 6,
      "phaseCycles": 9426,
      "phaseMs": 0.006284,
      "workCycles": 2521728,
      "averageActiveTiles": 267.528962444303,
      "usefulCycles": 720896.0,
      "estimatedWorkCycles": 2521728,
      "usefulUtilization": 0.28587381351200447,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05195619885792304,
      "meanCycles": 3582.0,
      "p50Cycles": 3582,
      "p95Cycles": 3582,
      "maximumCycles": 3582,
      "firstOffset": 185460,
      "lastOffset": 194886
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r9_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 8112,
      "phaseMs": 0.005408,
      "workCycles": 257856,
      "averageActiveTiles": 31.78698224852071,
      "usefulCycles": 10368.0,
      "estimatedWorkCycles": 257856,
      "usefulUtilization": 0.040208488458674606,
      "workLaneOccupancy": 0.5625,
      "deviceUsefulUtilization": 0.0008682788783123232,
      "meanCycles": 8058.0,
      "p50Cycles": 8058,
      "p95Cycles": 8058,
      "maximumCycles": 8058,
      "firstOffset": 20490,
      "lastOffset": 28602
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r16_p16_c72_p80",
      "dimensions": {},
      "sampleCount": 1248,
      "tileCount": 1248,
      "phaseCount": 6,
      "phaseCycles": 7626,
      "phaseMs": 0.005084,
      "workCycles": 2223936,
      "averageActiveTiles": 291.62549173878836,
      "usefulCycles": 718848.0,
      "estimatedWorkCycles": 2223936,
      "usefulUtilization": 0.32323232323232326,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06403721821229433,
      "meanCycles": 1782.0,
      "p50Cycles": 1782,
      "p95Cycles": 1782,
      "maximumCycles": 1782,
      "firstOffset": 183678,
      "lastOffset": 191304
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r16_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 1,
      "phaseCycles": 7266,
      "phaseMs": 0.004844,
      "workCycles": 10359360,
      "averageActiveTiles": 1425.7308009909166,
      "usefulCycles": 829440.0,
      "estimatedWorkCycles": 10359360,
      "usefulUtilization": 0.08006672226855713,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0775499946145837,
      "meanCycles": 7194.0,
      "p50Cycles": 7194,
      "p95Cycles": 7194,
      "maximumCycles": 7194,
      "firstOffset": 20472,
      "lastOffset": 27738
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 1280,
      "tileCount": 1280,
      "phaseCount": 39,
      "phaseCycles": 6546,
      "phaseMs": 0.004364,
      "workCycles": 1704960,
      "averageActiveTiles": 260.4582951420715,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1704960,
      "usefulUtilization": 0.015392736486486487,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.002723618140517276,
      "meanCycles": 1332.0,
      "p50Cycles": 1332,
      "p95Cycles": 1332,
      "maximumCycles": 1332,
      "firstOffset": 918918,
      "lastOffset": 925464
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r15_p15_c72_p80",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 5,
      "phaseCycles": 5730,
      "phaseMs": 0.00382,
      "workCycles": 399168,
      "averageActiveTiles": 69.6628272251309,
      "usefulCycles": 120960.0,
      "estimatedWorkCycles": 399168,
      "usefulUtilization": 0.30303030303030304,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.014340997040746642,
      "meanCycles": 1782.0,
      "p50Cycles": 1782,
      "p95Cycles": 1782,
      "maximumCycles": 1782,
      "firstOffset": 180054,
      "lastOffset": 187980
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c32_p32",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 3,
      "phaseCycles": 5340,
      "phaseMs": 0.00356,
      "workCycles": 261504,
      "averageActiveTiles": 48.970786516853934,
      "usefulCycles": 25600.0,
      "estimatedWorkCycles": 261504,
      "usefulUtilization": 0.09789525208027411,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0032567985670086306,
      "meanCycles": 4086.0,
      "p50Cycles": 4086,
      "p95Cycles": 4086,
      "maximumCycles": 4086,
      "firstOffset": 186846,
      "lastOffset": 192186
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 31,
      "phaseCycles": 4932,
      "phaseMs": 0.003288,
      "workCycles": 730944,
      "averageActiveTiles": 148.20437956204378,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 730944,
      "usefulUtilization": 0.028368794326241134,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.002856236115518883,
      "meanCycles": 846.0,
      "p50Cycles": 846,
      "p95Cycles": 846,
      "maximumCycles": 846,
      "firstOffset": 853506,
      "lastOffset": 863700
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 352,
      "tileCount": 352,
      "phaseCount": 5,
      "phaseCycles": 4512,
      "phaseMs": 0.003008,
      "workCycles": 540672,
      "averageActiveTiles": 119.82978723404256,
      "usefulCycles": 180224.0,
      "estimatedWorkCycles": 540672,
      "usefulUtilization": 0.3333333333333333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.027135368485969782,
      "meanCycles": 1536.0,
      "p50Cycles": 1536,
      "p95Cycles": 1536,
      "maximumCycles": 1536,
      "firstOffset": 188058,
      "lastOffset": 192570
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c224_p224",
      "dimensions": {},
      "sampleCount": 1368,
      "tileCount": 1368,
      "phaseCount": 51,
      "phaseCycles": 4164,
      "phaseMs": 0.002776,
      "workCycles": 5368032,
      "averageActiveTiles": 1289.1527377521613,
      "usefulCycles": 306432.0,
      "estimatedWorkCycles": 5368032,
      "usefulUtilization": 0.05708460754332314,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.04999373512091217,
      "meanCycles": 3924.0,
      "p50Cycles": 3924,
      "p95Cycles": 3924,
      "maximumCycles": 3924,
      "firstOffset": 961626,
      "lastOffset": 965790
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c48_p224",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 11,
      "phaseCycles": 3954,
      "phaseMs": 0.002636,
      "workCycles": 282528,
      "averageActiveTiles": 71.45371775417298,
      "usefulCycles": 3456.0,
      "estimatedWorkCycles": 282528,
      "usefulUtilization": 0.012232415902140673,
      "workLaneOccupancy": 0.10714285714285714,
      "deviceUsefulUtilization": 0.000593785049811968,
      "meanCycles": 3924.0,
      "p50Cycles": 3924,
      "p95Cycles": 3924,
      "maximumCycles": 3924,
      "firstOffset": 961296,
      "lastOffset": 965250
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c128_p128",
      "dimensions": {},
      "sampleCount": 1467,
      "tileCount": 1467,
      "phaseCount": 62,
      "phaseCycles": 2628,
      "phaseMs": 0.001752,
      "workCycles": 3644028,
      "averageActiveTiles": 1386.6164383561643,
      "usefulCycles": 187776.0,
      "estimatedWorkCycles": 3644028,
      "usefulUtilization": 0.05152979066022544,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.04854079809410363,
      "meanCycles": 2484.0,
      "p50Cycles": 2484,
      "p95Cycles": 2484,
      "maximumCycles": 2484,
      "firstOffset": 948930,
      "lastOffset": 951558
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 3,
      "phaseCycles": 2250,
      "phaseMs": 0.0015,
      "workCycles": 56640,
      "averageActiveTiles": 25.173333333333332,
      "usefulCycles": 6400.0,
      "estimatedWorkCycles": 56640,
      "usefulUtilization": 0.11299435028248588,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.001932367149758454,
      "meanCycles": 1770.0,
      "p50Cycles": 1770,
      "p95Cycles": 1770,
      "maximumCycles": 1770,
      "firstOffset": 187278,
      "lastOffset": 189528
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 59,
      "phaseCycles": 2100,
      "phaseMs": 0.0014,
      "workCycles": 2203200,
      "averageActiveTiles": 1049.142857142857,
      "usefulCycles": 77472.0,
      "estimatedWorkCycles": 2203200,
      "usefulUtilization": 0.03516339869281046,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.025062111801242237,
      "meanCycles": 1530.0,
      "p50Cycles": 1530,
      "p95Cycles": 1530,
      "maximumCycles": 1530,
      "firstOffset": 967512,
      "lastOffset": 969612
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 45,
      "phaseCycles": 2094,
      "phaseMs": 0.001396,
      "workCycles": 1975104,
      "averageActiveTiles": 943.2206303724928,
      "usefulCycles": 82944.0,
      "estimatedWorkCycles": 1975104,
      "usefulUtilization": 0.04199475065616798,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.026909181512395665,
      "meanCycles": 1524.0,
      "p50Cycles": 1524,
      "p95Cycles": 1524,
      "maximumCycles": 1524,
      "firstOffset": 934458,
      "lastOffset": 936552
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c32_r4_r4",
      "dimensions": {},
      "sampleCount": 954,
      "tileCount": 954,
      "phaseCount": 50,
      "phaseCycles": 1842,
      "phaseMs": 0.001228,
      "workCycles": 1625616,
      "averageActiveTiles": 882.527687296417,
      "usefulCycles": 61056.0,
      "estimatedWorkCycles": 1625616,
      "usefulUtilization": 0.03755868544600939,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.02251805693244583,
      "meanCycles": 1704.0,
      "p50Cycles": 1704,
      "p95Cycles": 1704,
      "maximumCycles": 1704,
      "firstOffset": 954210,
      "lastOffset": 956052
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1024,
      "tileCount": 1024,
      "phaseCount": 44,
      "phaseCycles": 1638,
      "phaseMs": 0.001092,
      "workCycles": 1218336,
      "averageActiveTiles": 743.7948717948718,
      "usefulCycles": 27976.0,
      "estimatedWorkCycles": 1218336,
      "usefulUtilization": 0.02296246684001786,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.011602829537612146,
      "meanCycles": 1189.78125,
      "p50Cycles": 1152,
      "p95Cycles": 1524,
      "maximumCycles": 1524,
      "firstOffset": 959028,
      "lastOffset": 960666
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 16,
      "phaseCycles": 1422,
      "phaseMs": 0.000948,
      "workCycles": 78336,
      "averageActiveTiles": 55.08860759493671,
      "usefulCycles": 4608.0,
      "estimatedWorkCycles": 78336,
      "usefulUtilization": 0.058823529411764705,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.002201430930104568,
      "meanCycles": 1224.0,
      "p50Cycles": 1224,
      "p95Cycles": 1224,
      "maximumCycles": 1224,
      "firstOffset": 943164,
      "lastOffset": 944586
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 45,
      "phaseCycles": 1242,
      "phaseMs": 0.000828,
      "workCycles": 870912,
      "averageActiveTiles": 701.2173913043479,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 870912,
      "usefulUtilization": 0.023809523809523808,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.011342155009451797,
      "meanCycles": 672.0,
      "p50Cycles": 672,
      "p95Cycles": 672,
      "maximumCycles": 672,
      "firstOffset": 938070,
      "lastOffset": 939312
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 513,
      "tileCount": 513,
      "phaseCount": 45,
      "phaseCycles": 1152,
      "phaseMs": 0.000768,
      "workCycles": 520182,
      "averageActiveTiles": 451.546875,
      "usefulCycles": 16416.0,
      "estimatedWorkCycles": 520182,
      "usefulUtilization": 0.03155818540433925,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.00968070652173913,
      "meanCycles": 1014.0,
      "p50Cycles": 1014,
      "p95Cycles": 1014,
      "maximumCycles": 1014,
      "firstOffset": 954216,
      "lastOffset": 955368
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p1_c72_p80",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 7,
      "phaseCycles": 852,
      "phaseMs": 0.000568,
      "workCycles": 24576,
      "averageActiveTiles": 28.845070422535212,
      "usefulCycles": 1152.0,
      "estimatedWorkCycles": 24576,
      "usefulUtilization": 0.046875,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0009185548071034905,
      "meanCycles": 768.0,
      "p50Cycles": 768,
      "p95Cycles": 768,
      "maximumCycles": 768,
      "firstOffset": 868584,
      "lastOffset": 869436
    }
  ],
  "samples": []
}
