{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 189315,
  "matchedSampleCount": 188516,
  "profileSpanCycles": 10666374,
  "profileSpanMs": 7.110916,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 10093932,
      "phaseMs": 6.729288,
      "workCycles": 14853856092,
      "averageActiveTiles": 1471.562924339098,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 10090934.845108695,
      "p50Cycles": 10090938,
      "p95Cycles": 10090950,
      "maximumCycles": 10090968,
      "firstOffset": 425952,
      "lastOffset": 10519884
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 18275,
      "tileCount": 1469,
      "phaseCount": 52,
      "phaseCycles": 458496,
      "phaseMs": 0.305664,
      "workCycles": 68219280,
      "averageActiveTiles": 148.78925879396985,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 3732.9291381668945,
      "p50Cycles": 2064,
      "p95Cycles": 12912,
      "maximumCycles": 48984,
      "firstOffset": 0,
      "lastOffset": 10664274
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 72989,
      "tileCount": 1472,
      "phaseCount": 62,
      "phaseCycles": 402432,
      "phaseMs": 0.268288,
      "workCycles": 409860918,
      "averageActiveTiles": 1018.4600578482824,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5615.3792763293095,
      "p50Cycles": 3846,
      "p95Cycles": 16908,
      "maximumCycles": 31830,
      "firstOffset": 0,
      "lastOffset": 10664856
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10600,
      "tileCount": 1472,
      "phaseCount": 133,
      "phaseCycles": 61584,
      "phaseMs": 0.041056,
      "workCycles": 42208656,
      "averageActiveTiles": 685.3834762275916,
      "usefulCycles": 23067536.0,
      "estimatedWorkCycles": 42208656,
      "usefulUtilization": 0.5465119761216751,
      "workLaneOccupancy": 0.9933504298002166,
      "deviceUsefulUtilization": 0.25446350407220397,
      "meanCycles": 3981.948679245283,
      "p50Cycles": 3546,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 19740,
      "lastOffset": 10662672
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 7911,
      "tileCount": 1472,
      "phaseCount": 147,
      "phaseCycles": 60276,
      "phaseMs": 0.040184,
      "workCycles": 7581294,
      "averageActiveTiles": 125.77632888711925,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 958.3230944254835,
      "p50Cycles": 1008,
      "p95Cycles": 1632,
      "maximumCycles": 1686,
      "firstOffset": 27144,
      "lastOffset": 10659798
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 25267,
      "tileCount": 1472,
      "phaseCount": 138,
      "phaseCycles": 41070,
      "phaseMs": 0.02738,
      "workCycles": 7966842,
      "averageActiveTiles": 193.98203067932798,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 5010318,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 315.30620968061106,
      "p50Cycles": 300,
      "p95Cycles": 564,
      "maximumCycles": 1260,
      "firstOffset": 6738,
      "lastOffset": 10658124
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 6722,
      "tileCount": 1458,
      "phaseCount": 122,
      "phaseCycles": 34080,
      "phaseMs": 0.02272,
      "workCycles": 25243782,
      "averageActiveTiles": 740.7213028169014,
      "usefulCycles": 3815754.0,
      "estimatedWorkCycles": 24871938,
      "usefulUtilization": 0.15341603054816236,
      "workLaneOccupancy": 0.9777025378858536,
      "deviceUsefulUtilization": 0.07606291622014696,
      "meanCycles": 3755.397500743826,
      "p50Cycles": 2088,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 0,
      "lastOffset": 10658106
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 30,
      "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": 276054,
      "lastOffset": 309576
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 22,
      "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": 276444,
      "lastOffset": 309054
    },
    {
      "name": "layer_norm_f8",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 34,
      "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": 10535136
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 32,
      "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": 369378,
      "lastOffset": 400188
    },
    {
      "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": 369402,
      "lastOffset": 400062
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 17,
      "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": 276060,
      "lastOffset": 306552
    },
    {
      "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": 276450,
      "lastOffset": 306096
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 19,
      "phaseCycles": 28500,
      "phaseMs": 0.019,
      "workCycles": 38108064,
      "averageActiveTiles": 1337.125052631579,
      "usefulCycles": 16866144.0,
      "estimatedWorkCycles": 38108064,
      "usefulUtilization": 0.4425872697180313,
      "workLaneOccupancy": 0.9972413793103448,
      "deviceUsefulUtilization": 0.4020343249427918,
      "meanCycles": 25888.630434782608,
      "p50Cycles": 25890,
      "p95Cycles": 25890,
      "maximumCycles": 25890,
      "firstOffset": 153600,
      "lastOffset": 182100
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3961,
      "tileCount": 1472,
      "phaseCount": 104,
      "phaseCycles": 22488,
      "phaseMs": 0.014992,
      "workCycles": 1632456,
      "averageActiveTiles": 72.5923159018143,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 412.1322898258016,
      "p50Cycles": 384,
      "p95Cycles": 522,
      "maximumCycles": 804,
      "firstOffset": 50310,
      "lastOffset": 10609812
    },
    {
      "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": "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": 10666374
    },
    {
      "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": 10545210,
      "lastOffset": 10562334
    },
    {
      "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": 10545384,
      "lastOffset": 10561938
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 11,
      "phaseCycles": 13086,
      "phaseMs": 0.008724,
      "workCycles": 11062656,
      "averageActiveTiles": 845.3810178817056,
      "usefulCycles": 6298560.0,
      "estimatedWorkCycles": 11062656,
      "usefulUtilization": 0.5693533270852859,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.3269840320554991,
      "meanCycles": 12804.0,
      "p50Cycles": 12804,
      "p95Cycles": 12804,
      "maximumCycles": 12804,
      "firstOffset": 134472,
      "lastOffset": 147558
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 12,
      "phaseCycles": 10902,
      "phaseMs": 0.007268,
      "workCycles": 5944320,
      "averageActiveTiles": 545.2504127682994,
      "usefulCycles": 3267378.0,
      "estimatedWorkCycles": 5944320,
      "usefulUtilization": 0.5496638808139535,
      "workLaneOccupancy": 0.875390625,
      "deviceUsefulUtilization": 0.20360357194132708,
      "meanCycles": 10320.0,
      "p50Cycles": 10320,
      "p95Cycles": 10320,
      "maximumCycles": 10320,
      "firstOffset": 133884,
      "lastOffset": 144786
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 950,
      "phaseCount": 29,
      "phaseCycles": 10470,
      "phaseMs": 0.00698,
      "workCycles": 4282080,
      "averageActiveTiles": 408.9856733524355,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0584693326689091,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 119640,
      "lastOffset": 10600986
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 20,
      "phaseCycles": 9960,
      "phaseMs": 0.00664,
      "workCycles": 9662976,
      "averageActiveTiles": 970.178313253012,
      "usefulCycles": 4408992.0,
      "estimatedWorkCycles": 9662976,
      "usefulUtilization": 0.4562768240343348,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3007268203247774,
      "meanCycles": 8388.0,
      "p50Cycles": 8388,
      "p95Cycles": 8388,
      "maximumCycles": 8388,
      "firstOffset": 189288,
      "lastOffset": 199248
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 2534,
      "tileCount": 1458,
      "phaseCount": 72,
      "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": 334476,
      "lastOffset": 10655292
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 150,
      "phaseCount": 21,
      "phaseCycles": 9222,
      "phaseMs": 0.006148,
      "workCycles": 314400,
      "averageActiveTiles": 34.09238776837996,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0023573119100826945,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 118758,
      "lastOffset": 10599588
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 96,
      "phaseCount": 6,
      "phaseCycles": 8964,
      "phaseMs": 0.005976,
      "workCycles": 809856,
      "averageActiveTiles": 90.34538152610442,
      "usefulCycles": 373977.0,
      "estimatedWorkCycles": 809856,
      "usefulUtilization": 0.46178209459459457,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.028342309455212152,
      "meanCycles": 8436.0,
      "p50Cycles": 8436,
      "p95Cycles": 8436,
      "maximumCycles": 8436,
      "firstOffset": 190338,
      "lastOffset": 199302
    },
    {
      "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": 10602894,
      "lastOffset": 10611072
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1694,
      "tileCount": 794,
      "phaseCount": 77,
      "phaseCycles": 8100,
      "phaseMs": 0.0054,
      "workCycles": 488592,
      "averageActiveTiles": 60.32,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 288.4250295159386,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 300,
      "firstOffset": 24420,
      "lastOffset": 10665006
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 29,
      "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": 224448,
      "lastOffset": 232146
    },
    {
      "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": 10619670,
      "lastOffset": 10627020
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 32,
      "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": 204234,
      "lastOffset": 211212
    },
    {
      "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": 10612230,
      "lastOffset": 10618350
    },
    {
      "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": 10649622,
      "lastOffset": 10653024
    },
    {
      "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": 10649136,
      "lastOffset": 10651716
    },
    {
      "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": 10641396,
      "lastOffset": 10643736
    },
    {
      "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": 10659186,
      "lastOffset": 10661412
    },
    {
      "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": 10644012,
      "lastOffset": 10646208
    },
    {
      "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": 10584000,
      "lastOffset": 10637094
    },
    {
      "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": 10629798,
      "lastOffset": 10631412
    }
  ],
  "samples": []
}
