{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 189647,
  "matchedSampleCount": 188848,
  "profileSpanCycles": 11081280,
  "profileSpanMs": 7.38752,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 10493448,
      "phaseMs": 6.995632,
      "workCycles": 15441943644,
      "averageActiveTiles": 1471.5795650771795,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 10490450.845108695,
      "p50Cycles": 10490454,
      "p95Cycles": 10490466,
      "maximumCycles": 10490484,
      "firstOffset": 441342,
      "lastOffset": 10934790
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 18275,
      "tileCount": 1469,
      "phaseCount": 52,
      "phaseCycles": 458964,
      "phaseMs": 0.305976,
      "workCycles": 68230362,
      "averageActiveTiles": 148.66168588386017,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 3733.5355403556773,
      "p50Cycles": 2064,
      "p95Cycles": 12930,
      "maximumCycles": 48984,
      "firstOffset": 0,
      "lastOffset": 11079180
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 72989,
      "tileCount": 1472,
      "phaseCount": 62,
      "phaseCycles": 417786,
      "phaseMs": 0.278524,
      "workCycles": 425880522,
      "averageActiveTiles": 1019.3748043256595,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5834.858978750222,
      "p50Cycles": 3942,
      "p95Cycles": 17442,
      "maximumCycles": 31830,
      "firstOffset": 0,
      "lastOffset": 11079762
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 7915,
      "tileCount": 1472,
      "phaseCount": 149,
      "phaseCycles": 62946,
      "phaseMs": 0.041964,
      "workCycles": 7584366,
      "averageActiveTiles": 120.49003908111715,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 958.2269109286166,
      "p50Cycles": 1008,
      "p95Cycles": 1632,
      "maximumCycles": 1686,
      "firstOffset": 27144,
      "lastOffset": 11074704
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10600,
      "tileCount": 1472,
      "phaseCount": 134,
      "phaseCycles": 61578,
      "phaseMs": 0.041052,
      "workCycles": 42208656,
      "averageActiveTiles": 685.4502582091006,
      "usefulCycles": 23067536.0,
      "estimatedWorkCycles": 42208656,
      "usefulUtilization": 0.5465119761216751,
      "workLaneOccupancy": 0.9933504298002166,
      "deviceUsefulUtilization": 0.2544882983335381,
      "meanCycles": 3981.948679245283,
      "p50Cycles": 3546,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 19740,
      "lastOffset": 11077578
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 25331,
      "tileCount": 1472,
      "phaseCount": 138,
      "phaseCycles": 48084,
      "phaseMs": 0.032056,
      "workCycles": 8309178,
      "averageActiveTiles": 172.8054654354879,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 5297742,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 328.0240811653705,
      "p50Cycles": 300,
      "p95Cycles": 720,
      "maximumCycles": 1260,
      "firstOffset": 6738,
      "lastOffset": 11073030
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 7010,
      "tileCount": 1458,
      "phaseCount": 129,
      "phaseCycles": 39840,
      "phaseMs": 0.02656,
      "workCycles": 31526214,
      "averageActiveTiles": 791.3206325301205,
      "usefulCycles": 5063802.0,
      "estimatedWorkCycles": 31154370,
      "usefulUtilization": 0.16253905952840644,
      "workLaneOccupancy": 0.9416934333283247,
      "deviceUsefulUtilization": 0.08634746185830278,
      "meanCycles": 4497.320114122682,
      "p50Cycles": 2088,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 0,
      "lastOffset": 11073012
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 31,
      "phaseCycles": 33522,
      "phaseMs": 0.022348,
      "workCycles": 23797800,
      "averageActiveTiles": 709.9158761410417,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23797800,
      "usefulUtilization": 0.6450344149459194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.31108707325234825,
      "meanCycles": 30510.0,
      "p50Cycles": 30510,
      "p95Cycles": 30510,
      "maximumCycles": 30510,
      "firstOffset": 291444,
      "lastOffset": 324966
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 23,
      "phaseCycles": 32610,
      "phaseMs": 0.02174,
      "workCycles": 18720000,
      "averageActiveTiles": 574.0570377184913,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18720000,
      "usefulUtilization": 0.638,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.24881004759809608,
      "meanCycles": 30000.0,
      "p50Cycles": 30000,
      "p95Cycles": 30000,
      "maximumCycles": 30000,
      "firstOffset": 291834,
      "lastOffset": 324444
    },
    {
      "name": "layer_norm_f8",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 35,
      "phaseCycles": 30894,
      "phaseMs": 0.020596,
      "workCycles": 20855232,
      "averageActiveTiles": 675.0576811031268,
      "usefulCycles": 8503056.0,
      "estimatedWorkCycles": 20855232,
      "usefulUtilization": 0.4077181208053691,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1869791094936121,
      "meanCycles": 9536.0,
      "p50Cycles": 9540,
      "p95Cycles": 9540,
      "maximumCycles": 9540,
      "firstOffset": 38298,
      "lastOffset": 10950042
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 33,
      "phaseCycles": 30810,
      "phaseMs": 0.02054,
      "workCycles": 32666490,
      "averageActiveTiles": 1060.2560856864654,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32666490,
      "usefulUtilization": 0.7233394221417728,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5210088480589307,
      "meanCycles": 26886.0,
      "p50Cycles": 26886,
      "p95Cycles": 26886,
      "maximumCycles": 26886,
      "firstOffset": 384768,
      "lastOffset": 415578
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 19,
      "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": 384792,
      "lastOffset": 415452
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 16,
      "phaseCycles": 30492,
      "phaseMs": 0.020328,
      "workCycles": 824940,
      "averageActiveTiles": 27.0543093270366,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 824940,
      "usefulUtilization": 0.6441195723325333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.011838457984703043,
      "meanCycles": 27498.0,
      "p50Cycles": 27498,
      "p95Cycles": 27498,
      "maximumCycles": 27498,
      "firstOffset": 291450,
      "lastOffset": 321942
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 14,
      "phaseCycles": 29646,
      "phaseMs": 0.019764,
      "workCycles": 649008,
      "averageActiveTiles": 21.891924711596843,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 649008,
      "usefulUtilization": 0.637009096960284,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.009473746733190782,
      "meanCycles": 27042.0,
      "p50Cycles": 27042,
      "p95Cycles": 27042,
      "maximumCycles": 27042,
      "firstOffset": 291840,
      "lastOffset": 321486
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 20,
      "phaseCycles": 28506,
      "phaseMs": 0.019004,
      "workCycles": 38108064,
      "averageActiveTiles": 1336.843611871185,
      "usefulCycles": 16866144.0,
      "estimatedWorkCycles": 38108064,
      "usefulUtilization": 0.4425872697180313,
      "workLaneOccupancy": 0.9972413793103448,
      "deviceUsefulUtilization": 0.40194970395248597,
      "meanCycles": 25888.630434782608,
      "p50Cycles": 25890,
      "p95Cycles": 25890,
      "maximumCycles": 25890,
      "firstOffset": 156978,
      "lastOffset": 185484
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3961,
      "tileCount": 1472,
      "phaseCount": 103,
      "phaseCycles": 22506,
      "phaseMs": 0.015004,
      "workCycles": 1632456,
      "averageActiveTiles": 72.53425753132498,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 412.1322898258016,
      "p50Cycles": 384,
      "p95Cycles": 522,
      "maximumCycles": 804,
      "firstOffset": 50310,
      "lastOffset": 11024718
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 3,
      "phaseCycles": 20304,
      "phaseMs": 0.013536,
      "workCycles": 18294984,
      "averageActiveTiles": 901.0531914893617,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18294984,
      "usefulUtilization": 0.8296674104771012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5078630897317299,
      "meanCycles": 18822.0,
      "p50Cycles": 18822,
      "p95Cycles": 18822,
      "maximumCycles": 18822,
      "firstOffset": 66960,
      "lastOffset": 87264
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 19302,
      "phaseMs": 0.012868,
      "workCycles": 9083340,
      "averageActiveTiles": 470.59061237177497,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 9083340,
      "usefulUtilization": 0.8252541466024612,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.2638293846548905,
      "meanCycles": 18690.0,
      "p50Cycles": 18690,
      "p95Cycles": 18690,
      "maximumCycles": 18690,
      "firstOffset": 67548,
      "lastOffset": 86850
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 12,
      "phaseCycles": 19236,
      "phaseMs": 0.012824,
      "workCycles": 11062656,
      "averageActiveTiles": 575.101684341859,
      "usefulCycles": 6298560.0,
      "estimatedWorkCycles": 11062656,
      "usefulUtilization": 0.5693533270852859,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.2224429737720036,
      "meanCycles": 12804.0,
      "p50Cycles": 12804,
      "p95Cycles": 12804,
      "maximumCycles": 12804,
      "firstOffset": 131724,
      "lastOffset": 150960
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11800,
      "tileCount": 1458,
      "phaseCount": 148,
      "phaseCycles": 18144,
      "phaseMs": 0.012096,
      "workCycles": 12935772,
      "averageActiveTiles": 712.9503968253969,
      "usefulCycles": 2310624.0,
      "estimatedWorkCycles": 12935772,
      "usefulUtilization": 0.17862281431676438,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.08651440648723258,
      "meanCycles": 1096.2518644067798,
      "p50Cycles": 942,
      "p95Cycles": 1806,
      "maximumCycles": 1806,
      "firstOffset": 24426,
      "lastOffset": 11081280
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 11,
      "phaseCycles": 17268,
      "phaseMs": 0.011512,
      "workCycles": 5944320,
      "averageActiveTiles": 344.23905489923555,
      "usefulCycles": 3267378.0,
      "estimatedWorkCycles": 5944320,
      "usefulUtilization": 0.5496638808139535,
      "workLaneOccupancy": 0.875390625,
      "deviceUsefulUtilization": 0.12854332530138685,
      "meanCycles": 10320.0,
      "p50Cycles": 10320,
      "p95Cycles": 10320,
      "maximumCycles": 10320,
      "firstOffset": 131118,
      "lastOffset": 148386
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 7,
      "phaseCycles": 17124,
      "phaseMs": 0.011416,
      "workCycles": 13369860,
      "averageActiveTiles": 780.7673440784863,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 13369860,
      "usefulUtilization": 0.6358899794014298,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.33728405593979466,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 10960116,
      "lastOffset": 10977240
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 7,
      "phaseCycles": 16554,
      "phaseMs": 0.011036,
      "workCycles": 10528704,
      "averageActiveTiles": 636.0217470097862,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10528704,
      "usefulUtilization": 0.6282619399310685,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.2714594134610839,
      "meanCycles": 16248.0,
      "p50Cycles": 16248,
      "p95Cycles": 16248,
      "maximumCycles": 16248,
      "firstOffset": 10960290,
      "lastOffset": 10976844
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 21,
      "phaseCycles": 12036,
      "phaseMs": 0.008024,
      "workCycles": 9662976,
      "averageActiveTiles": 802.839481555334,
      "usefulCycles": 4408992.0,
      "estimatedWorkCycles": 9662976,
      "usefulUtilization": 0.4562768240343348,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.2488566907971737,
      "meanCycles": 8388.0,
      "p50Cycles": 8388,
      "p95Cycles": 8388,
      "maximumCycles": 8388,
      "firstOffset": 202134,
      "lastOffset": 214170
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 31,
      "phaseCycles": 10980,
      "phaseMs": 0.00732,
      "workCycles": 4282080,
      "averageActiveTiles": 389.9890710382514,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.055753543993030806,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 118746,
      "lastOffset": 11015892
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 96,
      "phaseCount": 7,
      "phaseCycles": 10794,
      "phaseMs": 0.007196,
      "workCycles": 809856,
      "averageActiveTiles": 75.0283490828238,
      "usefulCycles": 373977.0,
      "estimatedWorkCycles": 809856,
      "usefulUtilization": 0.46178209459459457,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.023537193066196196,
      "meanCycles": 8436.0,
      "p50Cycles": 8436,
      "p95Cycles": 8436,
      "maximumCycles": 8436,
      "firstOffset": 203406,
      "lastOffset": 214200
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 2534,
      "tileCount": 1458,
      "phaseCount": 73,
      "phaseCycles": 9876,
      "phaseMs": 0.006584,
      "workCycles": 13055688,
      "averageActiveTiles": 1321.9611178614823,
      "usefulCycles": 8640280.0,
      "estimatedWorkCycles": 13055688,
      "usefulUtilization": 0.6618019670813212,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5943454267702115,
      "meanCycles": 5152.205209155485,
      "p50Cycles": 8556,
      "p95Cycles": 8556,
      "maximumCycles": 8556,
      "firstOffset": 349866,
      "lastOffset": 11070198
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 22,
      "phaseCycles": 8730,
      "phaseMs": 0.00582,
      "workCycles": 314400,
      "averageActiveTiles": 36.013745704467354,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.002490163852781513,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 118800,
      "lastOffset": 11014494
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 18,
      "phaseCycles": 8178,
      "phaseMs": 0.005452,
      "workCycles": 927360,
      "averageActiveTiles": 113.3969185619956,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 927360,
      "usefulUtilization": 0.014149844720496895,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0010900467319531724,
      "meanCycles": 1260.0,
      "p50Cycles": 1260,
      "p95Cycles": 1260,
      "maximumCycles": 1260,
      "firstOffset": 11017800,
      "lastOffset": 11025978
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 30,
      "phaseCycles": 7650,
      "phaseMs": 0.0051,
      "workCycles": 10437120,
      "averageActiveTiles": 1364.3294117647058,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10437120,
      "usefulUtilization": 0.7241721854304636,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.6712020460358056,
      "meanCycles": 7248.0,
      "p50Cycles": 7248,
      "p95Cycles": 7248,
      "maximumCycles": 7248,
      "firstOffset": 239874,
      "lastOffset": 247524
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k112_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 560,
      "tileCount": 560,
      "phaseCount": 22,
      "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": 11034576,
      "lastOffset": 11041926
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 34,
      "phaseCycles": 6978,
      "phaseMs": 0.004652,
      "workCycles": 4398336,
      "averageActiveTiles": 630.3147033533963,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2988.0,
      "p50Cycles": 2988,
      "p95Cycles": 2988,
      "maximumCycles": 2988,
      "firstOffset": 219120,
      "lastOffset": 226098
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1670,
      "tileCount": 797,
      "phaseCount": 82,
      "phaseCycles": 6204,
      "phaseMs": 0.004136,
      "workCycles": 484146,
      "averageActiveTiles": 78.03771760154739,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 289.9077844311377,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 492,
      "firstOffset": 24420,
      "lastOffset": 11079912
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 6120,
      "phaseMs": 0.00408,
      "workCycles": 97536,
      "averageActiveTiles": 15.937254901960785,
      "usefulCycles": 21696.0,
      "estimatedWorkCycles": 97536,
      "usefulUtilization": 0.22244094488188976,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.0024083546462063087,
      "meanCycles": 6096.0,
      "p50Cycles": 6096,
      "p95Cycles": 6096,
      "maximumCycles": 6096,
      "firstOffset": 11027136,
      "lastOffset": 11033256
    },
    {
      "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": 37,
      "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": 11064528,
      "lastOffset": 11067930
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 24,
      "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": 11064042,
      "lastOffset": 11066622
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 2340,
      "phaseMs": 0.00156,
      "workCycles": 6720,
      "averageActiveTiles": 2.871794871794872,
      "usefulCycles": 1008.0,
      "estimatedWorkCycles": 6720,
      "usefulUtilization": 0.15,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0002926421404682274,
      "meanCycles": 1680.0,
      "p50Cycles": 1680,
      "p95Cycles": 1680,
      "maximumCycles": 1680,
      "firstOffset": 11056302,
      "lastOffset": 11058642
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 38,
      "phaseCycles": 2226,
      "phaseMs": 0.001484,
      "workCycles": 2324160,
      "averageActiveTiles": 1044.0970350404314,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2324160,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.011821750849642564,
      "meanCycles": 1614.0,
      "p50Cycles": 1614,
      "p95Cycles": 1614,
      "maximumCycles": 1614,
      "firstOffset": 11074092,
      "lastOffset": 11076318
    },
    {
      "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": 11058918,
      "lastOffset": 11061114
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 46,
      "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": 10998906,
      "lastOffset": 11052000
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 1614,
      "phaseMs": 0.001076,
      "workCycles": 25536,
      "averageActiveTiles": 15.821561338289962,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1596.0,
      "p50Cycles": 1596,
      "p95Cycles": 1596,
      "maximumCycles": 1596,
      "firstOffset": 11044704,
      "lastOffset": 11046318
    }
  ],
  "samples": []
}
