{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 192529,
  "matchedSampleCount": 191730,
  "profileSpanCycles": 11016888,
  "profileSpanMs": 7.344592,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 9,
      "phaseCycles": 10429332,
      "phaseMs": 6.952888,
      "workCycles": 15347564892,
      "averageActiveTiles": 1471.5769803857045,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 10426334.845108695,
      "p50Cycles": 10426338,
      "p95Cycles": 10426350,
      "maximumCycles": 10426368,
      "firstOffset": 439686,
      "lastOffset": 10869018
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 18275,
      "tileCount": 1469,
      "phaseCount": 52,
      "phaseCycles": 461094,
      "phaseMs": 0.307396,
      "workCycles": 69709470,
      "averageActiveTiles": 151.18277401137294,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 3814.471682626539,
      "p50Cycles": 2064,
      "p95Cycles": 12912,
      "maximumCycles": 48984,
      "firstOffset": 0,
      "lastOffset": 11014788
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 72989,
      "tileCount": 1472,
      "phaseCount": 62,
      "phaseCycles": 418722,
      "phaseMs": 0.279148,
      "workCycles": 423051162,
      "averageActiveTiles": 1010.3389886368522,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5796.094781405417,
      "p50Cycles": 3708,
      "p95Cycles": 17154,
      "maximumCycles": 31830,
      "firstOffset": 0,
      "lastOffset": 11015370
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 9325,
      "tileCount": 1472,
      "phaseCount": 176,
      "phaseCycles": 71694,
      "phaseMs": 0.047796,
      "workCycles": 8886522,
      "averageActiveTiles": 123.9507071721483,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 952.9782305630026,
      "p50Cycles": 1008,
      "p95Cycles": 1632,
      "maximumCycles": 1686,
      "firstOffset": 19740,
      "lastOffset": 11012610
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10600,
      "tileCount": 1472,
      "phaseCount": 133,
      "phaseCycles": 61776,
      "phaseMs": 0.041184,
      "workCycles": 42453264,
      "averageActiveTiles": 687.2128982128983,
      "usefulCycles": 23280560.0,
      "estimatedWorkCycles": 42453264,
      "usefulUtilization": 0.5483809207226092,
      "workLaneOccupancy": 0.9925529621706577,
      "deviceUsefulUtilization": 0.25601524582502844,
      "meanCycles": 4005.0249056603775,
      "p50Cycles": 3546,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 20004,
      "lastOffset": 11013186
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 26371,
      "tileCount": 1472,
      "phaseCount": 147,
      "phaseCycles": 47928,
      "phaseMs": 0.031952,
      "workCycles": 8679834,
      "averageActiveTiles": 181.10152729093642,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 5372622,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 329.14314967198817,
      "p50Cycles": 300,
      "p95Cycles": 660,
      "maximumCycles": 1260,
      "firstOffset": 6738,
      "lastOffset": 11008368
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 23,
      "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": 289476,
      "lastOffset": 322998
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 27,
      "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": 289866,
      "lastOffset": 322476
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 28,
      "phaseCycles": 31722,
      "phaseMs": 0.021148,
      "workCycles": 33930912,
      "averageActiveTiles": 1069.6334405144694,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 33930912,
      "usefulUtilization": 0.4667743678684499,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.33918306592873415,
      "meanCycles": 22803.032258064515,
      "p50Cycles": 22986,
      "p95Cycles": 22986,
      "maximumCycles": 22986,
      "firstOffset": 157362,
      "lastOffset": 10967754
    },
    {
      "name": "layer_norm_f8",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 29,
      "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": 38682,
      "lastOffset": 10884270
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 25,
      "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": 383112,
      "lastOffset": 413922
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 24,
      "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": 383136,
      "lastOffset": 413796
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 14,
      "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": 289482,
      "lastOffset": 319974
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 13,
      "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": 289872,
      "lastOffset": 319518
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 5762,
      "tileCount": 1458,
      "phaseCount": 113,
      "phaseCycles": 28860,
      "phaseMs": 0.01924,
      "workCycles": 22957062,
      "averageActiveTiles": 795.4629937629937,
      "usefulCycles": 3699114.0,
      "estimatedWorkCycles": 22585218,
      "usefulUtilization": 0.16378473743312993,
      "workLaneOccupancy": 0.9786285125305035,
      "deviceUsefulUtilization": 0.08707501920817139,
      "meanCycles": 3984.2176327664006,
      "p50Cycles": 2088,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 0,
      "lastOffset": 11008350
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 5417,
      "tileCount": 1472,
      "phaseCount": 104,
      "phaseCycles": 24834,
      "phaseMs": 0.016556,
      "workCycles": 3178056,
      "averageActiveTiles": 127.97197390674076,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 586.6819272660144,
      "p50Cycles": 420,
      "p95Cycles": 1140,
      "maximumCycles": 1140,
      "firstOffset": 50694,
      "lastOffset": 10959216
    },
    {
      "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": 67344,
      "lastOffset": 87648
    },
    {
      "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": 67932,
      "lastOffset": 87234
    },
    {
      "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": 132108,
      "lastOffset": 151344
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11800,
      "tileCount": 1458,
      "phaseCount": 144,
      "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": 24816,
      "lastOffset": 11016888
    },
    {
      "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": 131502,
      "lastOffset": 148770
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 8,
      "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": 10894344,
      "lastOffset": 10911468
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k112_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1344,
      "tileCount": 1344,
      "phaseCount": 20,
      "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": 195102,
      "lastOffset": 211974
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 9,
      "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": 10894518,
      "lastOffset": 10911072
    },
    {
      "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": 196956,
      "lastOffset": 212136
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 32,
      "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": 119130,
      "lastOffset": 10950390
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 2534,
      "tileCount": 1458,
      "phaseCount": 69,
      "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": 348210,
      "lastOffset": 11005536
    },
    {
      "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": 119184,
      "lastOffset": 10948992
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 19,
      "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": 10952298,
      "lastOffset": 10960476
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 23,
      "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": 237354,
      "lastOffset": 245052
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 35,
      "phaseCycles": 7680,
      "phaseMs": 0.00512,
      "workCycles": 4423872,
      "averageActiveTiles": 576.025,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2973.032258064516,
      "p50Cycles": 2988,
      "p95Cycles": 2988,
      "maximumCycles": 2988,
      "firstOffset": 218058,
      "lastOffset": 10981104
    },
    {
      "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": 10969074,
      "lastOffset": 10976424
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1686,
      "tileCount": 797,
      "phaseCount": 85,
      "phaseCycles": 6510,
      "phaseMs": 0.00434,
      "workCycles": 488754,
      "averageActiveTiles": 75.07741935483871,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 289.88967971530246,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 492,
      "firstOffset": 24810,
      "lastOffset": 11015520
    },
    {
      "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": 38,
      "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": 10999584,
      "lastOffset": 11002986
    },
    {
      "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": 10999098,
      "lastOffset": 11001678
    },
    {
      "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": 10991358,
      "lastOffset": 10993698
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 40,
      "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": 11009430,
      "lastOffset": 11011656
    },
    {
      "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": 10993974,
      "lastOffset": 10996170
    },
    {
      "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": 10933134,
      "lastOffset": 10986786
    }
  ],
  "samples": []
}
