{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 214150,
  "matchedSampleCount": 213351,
  "profileSpanCycles": 11499450,
  "profileSpanMs": 7.6663,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 9,
      "phaseCycles": 10867380,
      "phaseMs": 7.24492,
      "workCycles": 15992371548,
      "averageActiveTiles": 1471.5940316801289,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 10864382.845108695,
      "p50Cycles": 10864386,
      "p95Cycles": 10864398,
      "maximumCycles": 10864416,
      "firstOffset": 458316,
      "lastOffset": 11325696
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 18289,
      "tileCount": 1469,
      "phaseCount": 53,
      "phaseCycles": 470016,
      "phaseMs": 0.313344,
      "workCycles": 77063478,
      "averageActiveTiles": 163.95926521650327,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 4213.65181256493,
      "p50Cycles": 2364,
      "p95Cycles": 12936,
      "maximumCycles": 40836,
      "firstOffset": 0,
      "lastOffset": 11497074
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 77391,
      "tileCount": 1472,
      "phaseCount": 65,
      "phaseCycles": 440184,
      "phaseMs": 0.293456,
      "workCycles": 434081742,
      "averageActiveTiles": 986.1370290605747,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5608.943443036012,
      "p50Cycles": 3732,
      "p95Cycles": 15480,
      "maximumCycles": 29634,
      "firstOffset": 0,
      "lastOffset": 11497656
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 20160,
      "tileCount": 1472,
      "phaseCount": 206,
      "phaseCycles": 84432,
      "phaseMs": 0.056288,
      "workCycles": 16714542,
      "averageActiveTiles": 197.96453951108583,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 829.0943452380952,
      "p50Cycles": 594,
      "p95Cycles": 1584,
      "maximumCycles": 1686,
      "firstOffset": 19740,
      "lastOffset": 11495478
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10600,
      "tileCount": 1472,
      "phaseCount": 141,
      "phaseCycles": 54192,
      "phaseMs": 0.036128,
      "workCycles": 42453264,
      "averageActiveTiles": 783.386182462356,
      "usefulCycles": 23455232.0,
      "estimatedWorkCycles": 42453264,
      "usefulUtilization": 0.5524953746783757,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.29403345271562625,
      "meanCycles": 4005.0249056603775,
      "p50Cycles": 3546,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 20004,
      "lastOffset": 11495208
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 8861,
      "tileCount": 1458,
      "phaseCount": 131,
      "phaseCycles": 52662,
      "phaseMs": 0.035108,
      "workCycles": 46462758,
      "averageActiveTiles": 882.2824427480916,
      "usefulCycles": 8537418.0,
      "estimatedWorkCycles": 46168062,
      "usefulUtilization": 0.18492043265753716,
      "workLaneOccupancy": 0.9898455652173913,
      "deviceUsefulUtilization": 0.11013399336705124,
      "meanCycles": 5243.511793251326,
      "p50Cycles": 4830,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 0,
      "lastOffset": 11490228
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 22671,
      "tileCount": 1472,
      "phaseCount": 130,
      "phaseCycles": 47886,
      "phaseMs": 0.031924,
      "workCycles": 7900242,
      "averageActiveTiles": 164.98020298208243,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 4551366,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 348.47346830752946,
      "p50Cycles": 306,
      "p95Cycles": 660,
      "maximumCycles": 1500,
      "firstOffset": 6738,
      "lastOffset": 11490396
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 28,
      "phaseCycles": 32898,
      "phaseMs": 0.021932,
      "workCycles": 34993632,
      "averageActiveTiles": 1063.7008936713478,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 34993632,
      "usefulUtilization": 0.4525989185689556,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.32705833842152426,
      "meanCycles": 23517.225806451614,
      "p50Cycles": 23706,
      "p95Cycles": 23706,
      "maximumCycles": 23706,
      "firstOffset": 163434,
      "lastOffset": 11444442
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 21,
      "phaseCycles": 32790,
      "phaseMs": 0.02186,
      "workCycles": 23797800,
      "averageActiveTiles": 725.7639524245196,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23797800,
      "usefulUtilization": 0.6450344149459194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.318031743506106,
      "meanCycles": 30510.0,
      "p50Cycles": 30510,
      "p95Cycles": 30510,
      "maximumCycles": 30510,
      "firstOffset": 303918,
      "lastOffset": 336708
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 20,
      "phaseCycles": 32268,
      "phaseMs": 0.021512,
      "workCycles": 18720000,
      "averageActiveTiles": 580.1413164745259,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18720000,
      "usefulUtilization": 0.638,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.25144711950458387,
      "meanCycles": 30000.0,
      "p50Cycles": 30000,
      "p95Cycles": 30000,
      "maximumCycles": 30000,
      "firstOffset": 303930,
      "lastOffset": 336198
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 22,
      "phaseCycles": 30804,
      "phaseMs": 0.020536,
      "workCycles": 32666490,
      "averageActiveTiles": 1060.4626022594468,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32666490,
      "usefulUtilization": 0.7233394221417728,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5211103301095849,
      "meanCycles": 26886.0,
      "p50Cycles": 26886,
      "p95Cycles": 26886,
      "maximumCycles": 26886,
      "firstOffset": 404628,
      "lastOffset": 435432
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 18,
      "phaseCycles": 30660,
      "phaseMs": 0.02044,
      "workCycles": 6508512,
      "averageActiveTiles": 212.28023483365948,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6508512,
      "usefulUtilization": 0.7082393026240099,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.10213668850506254,
      "meanCycles": 26784.0,
      "p50Cycles": 26784,
      "p95Cycles": 26784,
      "maximumCycles": 26784,
      "firstOffset": 404646,
      "lastOffset": 435306
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 11,
      "phaseCycles": 28434,
      "phaseMs": 0.018956,
      "workCycles": 649008,
      "averageActiveTiles": 22.82506857986917,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 649008,
      "usefulUtilization": 0.637009096960284,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.009877565437580851,
      "meanCycles": 27042.0,
      "p50Cycles": 27042,
      "p95Cycles": 27042,
      "maximumCycles": 27042,
      "firstOffset": 304716,
      "lastOffset": 333150
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 13,
      "phaseCycles": 28224,
      "phaseMs": 0.018816,
      "workCycles": 824940,
      "averageActiveTiles": 29.228316326530614,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 824940,
      "usefulUtilization": 0.6441195723325333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.01278976264418811,
      "meanCycles": 27498.0,
      "p50Cycles": 27498,
      "p95Cycles": 27498,
      "maximumCycles": 27498,
      "firstOffset": 304728,
      "lastOffset": 332952
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 5400,
      "tileCount": 1472,
      "phaseCount": 105,
      "phaseCycles": 23796,
      "phaseMs": 0.015864,
      "workCycles": 3171156,
      "averageActiveTiles": 133.26424609178014,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 587.2511111111111,
      "p50Cycles": 420,
      "p95Cycles": 1140,
      "maximumCycles": 1140,
      "firstOffset": 112134,
      "lastOffset": 11435724
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 29,
      "phaseCycles": 23718,
      "phaseMs": 0.015812,
      "workCycles": 15623928,
      "averageActiveTiles": 658.7371616493803,
      "usefulCycles": 5983632.0,
      "estimatedWorkCycles": 15623928,
      "usefulUtilization": 0.3829787234042553,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.17138744376862922,
      "meanCycles": 7144.0,
      "p50Cycles": 7146,
      "p95Cycles": 7146,
      "maximumCycles": 7146,
      "firstOffset": 39834,
      "lastOffset": 11338560
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 13258,
      "tileCount": 1458,
      "phaseCount": 158,
      "phaseCycles": 23232,
      "phaseMs": 0.015488,
      "workCycles": 18984480,
      "averageActiveTiles": 817.1694214876034,
      "usefulCycles": 3095028.0,
      "estimatedWorkCycles": 18984480,
      "usefulUtilization": 0.16302937978812165,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.09050449997754222,
      "meanCycles": 1431.9263840699955,
      "p50Cycles": 1074,
      "p95Cycles": 2658,
      "maximumCycles": 2658,
      "firstOffset": 25500,
      "lastOffset": 11499450
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 12,
      "phaseCycles": 19488,
      "phaseMs": 0.012992,
      "workCycles": 11062656,
      "averageActiveTiles": 567.6650246305419,
      "usefulCycles": 6298560.0,
      "estimatedWorkCycles": 11062656,
      "usefulUtilization": 0.5693533270852859,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.21956655600771044,
      "meanCycles": 12804.0,
      "p50Cycles": 12804,
      "p95Cycles": 12804,
      "maximumCycles": 12804,
      "firstOffset": 138144,
      "lastOffset": 157632
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 4,
      "phaseCycles": 19266,
      "phaseMs": 0.012844,
      "workCycles": 18294984,
      "averageActiveTiles": 949.5995017128621,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18294984,
      "usefulUtilization": 0.8296674104771012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5352253801470489,
      "meanCycles": 18822.0,
      "p50Cycles": 18822,
      "p95Cycles": 18822,
      "maximumCycles": 18822,
      "firstOffset": 71670,
      "lastOffset": 90936
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 18990,
      "phaseMs": 0.01266,
      "workCycles": 9083340,
      "averageActiveTiles": 478.3222748815166,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 9083340,
      "usefulUtilization": 0.8252541466024612,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.2681640222542757,
      "meanCycles": 18690.0,
      "p50Cycles": 18690,
      "p95Cycles": 18690,
      "maximumCycles": 18690,
      "firstOffset": 71574,
      "lastOffset": 90564
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 13,
      "phaseCycles": 17460,
      "phaseMs": 0.01164,
      "workCycles": 5944320,
      "averageActiveTiles": 340.45360824742266,
      "usefulCycles": 3267378.0,
      "estimatedWorkCycles": 5944320,
      "usefulUtilization": 0.5496638808139535,
      "workLaneOccupancy": 0.875390625,
      "deviceUsefulUtilization": 0.12712979045271178,
      "meanCycles": 10320.0,
      "p50Cycles": 10320,
      "p95Cycles": 10320,
      "maximumCycles": 10320,
      "firstOffset": 137532,
      "lastOffset": 154992
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k112_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1344,
      "tileCount": 1344,
      "phaseCount": 19,
      "phaseCycles": 16872,
      "phaseMs": 0.011248,
      "workCycles": 19184256,
      "averageActiveTiles": 1137.0469416785206,
      "usefulCycles": 8817984.0,
      "estimatedWorkCycles": 19184256,
      "usefulUtilization": 0.45964691046658257,
      "workLaneOccupancy": 0.8368622448979591,
      "deviceUsefulUtilization": 0.3550544251345167,
      "meanCycles": 14274.0,
      "p50Cycles": 14274,
      "p95Cycles": 14274,
      "maximumCycles": 14274,
      "firstOffset": 202104,
      "lastOffset": 218976
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 9,
      "phaseCycles": 16620,
      "phaseMs": 0.01108,
      "workCycles": 13369860,
      "averageActiveTiles": 804.4440433212997,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 13369860,
      "usefulUtilization": 0.6358899794014298,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.34751216449536965,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 11357616,
      "lastOffset": 11374236
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 8,
      "phaseCycles": 16332,
      "phaseMs": 0.010888,
      "workCycles": 10528704,
      "averageActiveTiles": 644.6671565025716,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10528704,
      "usefulUtilization": 0.6282619399310685,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.27514934670798324,
      "meanCycles": 16248.0,
      "p50Cycles": 16248,
      "p95Cycles": 16248,
      "maximumCycles": 16248,
      "firstOffset": 11357622,
      "lastOffset": 11373954
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 32,
      "phaseCycles": 15576,
      "phaseMs": 0.010384,
      "workCycles": 4282080,
      "averageActiveTiles": 274.91525423728814,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.03930238270695161,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 125760,
      "lastOffset": 11426952
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k112_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 7,
      "phaseCycles": 15180,
      "phaseMs": 0.01012,
      "workCycles": 1608768,
      "averageActiveTiles": 105.97944664031621,
      "usefulCycles": 747954.0,
      "estimatedWorkCycles": 1608768,
      "usefulUtilization": 0.4649234693877551,
      "workLaneOccupancy": 0.8368622448979591,
      "deviceUsefulUtilization": 0.03347305164117546,
      "meanCycles": 14364.0,
      "p50Cycles": 14364,
      "p95Cycles": 14364,
      "maximumCycles": 14364,
      "firstOffset": 203958,
      "lastOffset": 219138
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 27,
      "phaseCycles": 14328,
      "phaseMs": 0.009552,
      "workCycles": 16908144,
      "averageActiveTiles": 1180.077051926298,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 16908144,
      "usefulUtilization": 0.556704981930601,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4463007974655888,
      "meanCycles": 11486.510869565218,
      "p50Cycles": 11376,
      "p95Cycles": 11880,
      "maximumCycles": 12360,
      "firstOffset": 364548,
      "lastOffset": 378876
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 22,
      "phaseCycles": 13356,
      "phaseMs": 0.008904,
      "workCycles": 314400,
      "averageActiveTiles": 23.539982030548067,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0016276677474380509,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 125832,
      "lastOffset": 11426130
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 14,
      "phaseCycles": 13164,
      "phaseMs": 0.008776,
      "workCycles": 2072928,
      "averageActiveTiles": 157.46946216955334,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0232518198512412,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 11411358,
      "lastOffset": 11424522
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 6,
      "phaseCycles": 9576,
      "phaseMs": 0.006384,
      "workCycles": 146592,
      "averageActiveTiles": 15.308270676691729,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.001135084087029167,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 11411934,
      "lastOffset": 11421510
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1826,
      "tileCount": 815,
      "phaseCount": 108,
      "phaseCycles": 9204,
      "phaseMs": 0.006136,
      "workCycles": 523134,
      "averageActiveTiles": 56.837679269882656,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 286.4917853231106,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 1068,
      "firstOffset": 25494,
      "lastOffset": 11497806
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 19,
      "phaseCycles": 8118,
      "phaseMs": 0.005412,
      "workCycles": 927360,
      "averageActiveTiles": 114.2350332594235,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 927360,
      "usefulUtilization": 0.014149844720496895,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0010981032488190494,
      "meanCycles": 1260.0,
      "p50Cycles": 1260,
      "p95Cycles": 1260,
      "maximumCycles": 1260,
      "firstOffset": 11428866,
      "lastOffset": 11436984
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 21,
      "phaseCycles": 7698,
      "phaseMs": 0.005132,
      "workCycles": 10437120,
      "averageActiveTiles": 1355.8222915042868,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10437120,
      "usefulUtilization": 0.7241721854304636,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.6670168423192924,
      "meanCycles": 7248.0,
      "p50Cycles": 7248,
      "p95Cycles": 7248,
      "maximumCycles": 7248,
      "firstOffset": 244356,
      "lastOffset": 252054
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k112_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 560,
      "tileCount": 560,
      "phaseCount": 23,
      "phaseCycles": 7350,
      "phaseMs": 0.0049,
      "workCycles": 900480,
      "averageActiveTiles": 122.51428571428572,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 900480,
      "usefulUtilization": 0.01457222814498934,
      "workLaneOccupancy": 0.8368622448979591,
      "deviceUsefulUtilization": 0.001212843833185448,
      "meanCycles": 1608.0,
      "p50Cycles": 1608,
      "p95Cycles": 1608,
      "maximumCycles": 1608,
      "firstOffset": 11445762,
      "lastOffset": 11453112
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 34,
      "phaseCycles": 7140,
      "phaseMs": 0.00476,
      "workCycles": 4423872,
      "averageActiveTiles": 619.5899159663866,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2973.032258064516,
      "p50Cycles": 2988,
      "p95Cycles": 2988,
      "maximumCycles": 2988,
      "firstOffset": 225600,
      "lastOffset": 11458860
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 3,
      "phaseCycles": 4404,
      "phaseMs": 0.002936,
      "workCycles": 1458000,
      "averageActiveTiles": 331.06267029972753,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1458000,
      "usefulUtilization": 0.6678518518518518,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.15020435967302453,
      "meanCycles": 4050.0,
      "p50Cycles": 4050,
      "p95Cycles": 4050,
      "maximumCycles": 4050,
      "firstOffset": 12372,
      "lastOffset": 16776
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 2,
      "phaseCycles": 4350,
      "phaseMs": 0.0029,
      "workCycles": 4348080,
      "averageActiveTiles": 999.5586206896552,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4348080,
      "usefulUtilization": 0.6633134624937904,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.45042166416791607,
      "meanCycles": 4026.0,
      "p50Cycles": 4026,
      "p95Cycles": 4026,
      "maximumCycles": 4026,
      "firstOffset": 12378,
      "lastOffset": 16728
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c32_r2_r2",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 41,
      "phaseCycles": 3402,
      "phaseMs": 0.002268,
      "workCycles": 1456560,
      "averageActiveTiles": 428.14814814814815,
      "usefulCycles": 24480.0,
      "estimatedWorkCycles": 1456560,
      "usefulUtilization": 0.01680672268907563,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.004888428801472279,
      "meanCycles": 2142.0,
      "p50Cycles": 2142,
      "p95Cycles": 2142,
      "maximumCycles": 2142,
      "firstOffset": 11478798,
      "lastOffset": 11482200
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 9,
      "phaseCycles": 3150,
      "phaseMs": 0.0021,
      "workCycles": 142080,
      "averageActiveTiles": 45.10476190476191,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 142080,
      "usefulUtilization": 0.11261261261261261,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.003450655624568668,
      "meanCycles": 1776.0,
      "p50Cycles": 1776,
      "p95Cycles": 1776,
      "maximumCycles": 1776,
      "firstOffset": 124038,
      "lastOffset": 127188
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 14,
      "phaseCycles": 2928,
      "phaseMs": 0.001952,
      "workCycles": 1729920,
      "averageActiveTiles": 590.8196721311475,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1729920,
      "usefulUtilization": 0.22641509433962265,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.09087669280114041,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 122640,
      "lastOffset": 125568
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 4,
      "phaseCycles": 2916,
      "phaseMs": 0.001944,
      "workCycles": 142464,
      "averageActiveTiles": 48.8559670781893,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 142464,
      "usefulUtilization": 0.19811320754716982,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0065754159957058505,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 122424,
      "lastOffset": 125340
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 13,
      "phaseCycles": 2892,
      "phaseMs": 0.001928,
      "workCycles": 1356960,
      "averageActiveTiles": 469.2116182572614,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1356960,
      "usefulUtilization": 0.3320363164721141,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10583919658428047,
      "meanCycles": 1542.0,
      "p50Cycles": 1542,
      "p95Cycles": 1542,
      "maximumCycles": 1542,
      "firstOffset": 124218,
      "lastOffset": 127110
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 28,
      "phaseCycles": 2580,
      "phaseMs": 0.00172,
      "workCycles": 1002672,
      "averageActiveTiles": 388.63255813953486,
      "usefulCycles": 14256.0,
      "estimatedWorkCycles": 1002672,
      "usefulUtilization": 0.014218009478672985,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.0037537917087967646,
      "meanCycles": 1266.0,
      "p50Cycles": 1266,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 11478312,
      "lastOffset": 11480892
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 2424,
      "phaseMs": 0.001616,
      "workCycles": 7056,
      "averageActiveTiles": 2.910891089108911,
      "usefulCycles": 1008.0,
      "estimatedWorkCycles": 7056,
      "usefulUtilization": 0.14285714285714285,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.000282501076194576,
      "meanCycles": 1764.0,
      "p50Cycles": 1764,
      "p95Cycles": 1764,
      "maximumCycles": 1764,
      "firstOffset": 11470488,
      "lastOffset": 11472912
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 46,
      "phaseCycles": 2232,
      "phaseMs": 0.001488,
      "workCycles": 2324160,
      "averageActiveTiles": 1041.2903225806451,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2324160,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.011789971949509117,
      "meanCycles": 1614.0,
      "p50Cycles": 1614,
      "p95Cycles": 1614,
      "maximumCycles": 1614,
      "firstOffset": 11491452,
      "lastOffset": 11493684
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 2196,
      "phaseMs": 0.001464,
      "workCycles": 8784,
      "averageActiveTiles": 4.0,
      "usefulCycles": 1728.0,
      "estimatedWorkCycles": 8784,
      "usefulUtilization": 0.19672131147540983,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0005345687811831789,
      "meanCycles": 2196.0,
      "p50Cycles": 2196,
      "p95Cycles": 2196,
      "maximumCycles": 2196,
      "firstOffset": 11473188,
      "lastOffset": 11475384
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c224_p224",
      "dimensions": {},
      "sampleCount": 1368,
      "tileCount": 1368,
      "phaseCount": 46,
      "phaseCycles": 2184,
      "phaseMs": 0.001456,
      "workCycles": 2766096,
      "averageActiveTiles": 1266.5274725274726,
      "usefulCycles": 153216.0,
      "estimatedWorkCycles": 2766096,
      "usefulUtilization": 0.05539070227497527,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.047658862876254184,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 11486040,
      "lastOffset": 11488224
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c48_p224",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 7,
      "phaseCycles": 2046,
      "phaseMs": 0.001364,
      "workCycles": 145584,
      "averageActiveTiles": 71.15542521994135,
      "usefulCycles": 1728.0,
      "estimatedWorkCycles": 145584,
      "usefulUtilization": 0.011869436201780416,
      "workLaneOccupancy": 0.10714285714285714,
      "deviceUsefulUtilization": 0.000573760040800714,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 11485788,
      "lastOffset": 11487834
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 48,
      "phaseCycles": 1692,
      "phaseMs": 0.001128,
      "workCycles": 1928448,
      "averageActiveTiles": 1139.7446808510638,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 1928448,
      "usefulUtilization": 0.010752688172043012,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.008325624421831638,
      "meanCycles": 744.0,
      "p50Cycles": 744,
      "p95Cycles": 744,
      "maximumCycles": 744,
      "firstOffset": 11396772,
      "lastOffset": 11465652
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 34,
      "phaseCycles": 1410,
      "phaseMs": 0.00094,
      "workCycles": 1065312,
      "averageActiveTiles": 755.5404255319149,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1065312,
      "usefulUtilization": 0.038929440389294405,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.01998149861239593,
      "meanCycles": 822.0,
      "p50Cycles": 822,
      "p95Cycles": 822,
      "maximumCycles": 822,
      "firstOffset": 11460576,
      "lastOffset": 11461986
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p1_c72_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 5,
      "phaseCycles": 1320,
      "phaseMs": 0.00088,
      "workCycles": 12288,
      "averageActiveTiles": 9.309090909090909,
      "usefulCycles": 576.0,
      "estimatedWorkCycles": 12288,
      "usefulUtilization": 0.046875,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0002964426877470356,
      "meanCycles": 768.0,
      "p50Cycles": 768,
      "p95Cycles": 768,
      "maximumCycles": 768,
      "firstOffset": 11403192,
      "lastOffset": 11404512
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1076,
      "tileCount": 1076,
      "phaseCount": 40,
      "phaseCycles": 1020,
      "phaseMs": 0.00068,
      "workCycles": 839280,
      "averageActiveTiles": 822.8235294117648,
      "usefulCycles": 13988.0,
      "estimatedWorkCycles": 839280,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.009316389599317989,
      "meanCycles": 780.0,
      "p50Cycles": 780,
      "p95Cycles": 780,
      "maximumCycles": 780,
      "firstOffset": 11484252,
      "lastOffset": 11485272
    }
  ],
  "samples": []
}
