{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 241784,
  "matchedSampleCount": 241784,
  "profileSpanCycles": 12948846,
  "profileSpanMs": 8.632564,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 11,
      "phaseCycles": 12200724,
      "phaseMs": 8.133816,
      "workCycles": 17955046632,
      "averageActiveTiles": 1471.637800510855,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 12197721.89673913,
      "p50Cycles": 12197724,
      "p95Cycles": 12197736,
      "maximumCycles": 12197760,
      "firstOffset": 527754,
      "lastOffset": 12728478
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 25121,
      "tileCount": 1457,
      "phaseCount": 63,
      "phaseCycles": 592728,
      "phaseMs": 0.395152,
      "workCycles": 128954034,
      "averageActiveTiles": 217.5602198647609,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5133.316110027467,
      "p50Cycles": 2202,
      "p95Cycles": 19674,
      "maximumCycles": 46848,
      "firstOffset": 0,
      "lastOffset": 12948120
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 82335,
      "tileCount": 1472,
      "phaseCount": 73,
      "phaseCycles": 520824,
      "phaseMs": 0.347216,
      "workCycles": 512043168,
      "averageActiveTiles": 983.1405004377679,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6219.021898342139,
      "p50Cycles": 4470,
      "p95Cycles": 16518,
      "maximumCycles": 34158,
      "firstOffset": 0,
      "lastOffset": 12948696
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 29419,
      "tileCount": 1472,
      "phaseCount": 235,
      "phaseCycles": 118338,
      "phaseMs": 0.078892,
      "workCycles": 23872944,
      "averageActiveTiles": 201.73523297672767,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 811.4804718039362,
      "p50Cycles": 576,
      "p95Cycles": 1572,
      "maximumCycles": 10578,
      "firstOffset": 38106,
      "lastOffset": 12944880
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 16004,
      "tileCount": 1458,
      "phaseCount": 176,
      "phaseCycles": 91560,
      "phaseMs": 0.06104,
      "workCycles": 73422492,
      "averageActiveTiles": 801.9057667103539,
      "usefulCycles": 13615968.0,
      "estimatedWorkCycles": 73422492,
      "usefulUtilization": 0.1854468246596697,
      "workLaneOccupancy": 0.9656501217845392,
      "deviceUsefulUtilization": 0.10102641176135392,
      "meanCycles": 4587.758810297426,
      "p50Cycles": 3810,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 18306,
      "lastOffset": 12935232
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10440,
      "tileCount": 1472,
      "phaseCount": 147,
      "phaseCycles": 63216,
      "phaseMs": 0.042144,
      "workCycles": 42675360,
      "averageActiveTiles": 675.0721336370539,
      "usefulCycles": 23746848.0,
      "estimatedWorkCycles": 42675360,
      "usefulUtilization": 0.5564533726253276,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2551944059291539,
      "meanCycles": 4087.67816091954,
      "p50Cycles": 3690,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 38370,
      "lastOffset": 12944610
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 16172,
      "tileCount": 1472,
      "phaseCount": 112,
      "phaseCycles": 62472,
      "phaseMs": 0.041648,
      "workCycles": 7283328,
      "averageActiveTiles": 116.58547829427583,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 3429072,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 450.36655948553056,
      "p50Cycles": 360,
      "p95Cycles": 1410,
      "maximumCycles": 3150,
      "firstOffset": 28926,
      "lastOffset": 12935850
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 5905,
      "tileCount": 1472,
      "phaseCount": 132,
      "phaseCycles": 54492,
      "phaseMs": 0.036328,
      "workCycles": 3002328,
      "averageActiveTiles": 55.096674741246424,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 508.43827265029637,
      "p50Cycles": 420,
      "p95Cycles": 720,
      "maximumCycles": 5430,
      "firstOffset": 0,
      "lastOffset": 12941514
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 23,
      "phaseCycles": 34482,
      "phaseMs": 0.022988,
      "workCycles": 23797800,
      "averageActiveTiles": 690.1513833304333,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23797800,
      "usefulUtilization": 0.6450344149459194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3024262185942004,
      "meanCycles": 30510.0,
      "p50Cycles": 30510,
      "p95Cycles": 30510,
      "maximumCycles": 30510,
      "firstOffset": 347868,
      "lastOffset": 382350
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 24,
      "phaseCycles": 33942,
      "phaseMs": 0.022628,
      "workCycles": 18720000,
      "averageActiveTiles": 551.5290790171468,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18720000,
      "usefulUtilization": 0.638,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.23904589158487752,
      "meanCycles": 30000.0,
      "p50Cycles": 30000,
      "p95Cycles": 30000,
      "maximumCycles": 30000,
      "firstOffset": 347916,
      "lastOffset": 381858
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 28,
      "phaseCycles": 32736,
      "phaseMs": 0.021824,
      "workCycles": 32666490,
      "averageActiveTiles": 997.8766495601174,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32666490,
      "usefulUtilization": 0.7233394221417728,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.49035565153640187,
      "meanCycles": 26886.0,
      "p50Cycles": 26886,
      "p95Cycles": 26886,
      "maximumCycles": 26886,
      "firstOffset": 467550,
      "lastOffset": 500286
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 19,
      "phaseCycles": 32556,
      "phaseMs": 0.021704,
      "workCycles": 6508512,
      "averageActiveTiles": 199.91743457427202,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6508512,
      "usefulUtilization": 0.7082393026240099,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.09618844051988013,
      "meanCycles": 26784.0,
      "p50Cycles": 26784,
      "p95Cycles": 26784,
      "maximumCycles": 26784,
      "firstOffset": 467616,
      "lastOffset": 500172
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 27,
      "phaseCycles": 32538,
      "phaseMs": 0.021692,
      "workCycles": 34993632,
      "averageActiveTiles": 1075.4696662364006,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 34993632,
      "usefulUtilization": 0.4525989185689556,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.330676907535537,
      "meanCycles": 23517.225806451614,
      "p50Cycles": 23706,
      "p95Cycles": 23706,
      "maximumCycles": 23706,
      "firstOffset": 193338,
      "lastOffset": 12868536
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 13,
      "phaseCycles": 29304,
      "phaseMs": 0.019536,
      "workCycles": 824940,
      "averageActiveTiles": 28.151105651105652,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 824940,
      "usefulUtilization": 0.6441195723325333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.012318395470569383,
      "meanCycles": 27498.0,
      "p50Cycles": 27498,
      "p95Cycles": 27498,
      "maximumCycles": 27498,
      "firstOffset": 349200,
      "lastOffset": 378504
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 13,
      "phaseCycles": 28770,
      "phaseMs": 0.01918,
      "workCycles": 649008,
      "averageActiveTiles": 22.5584984358707,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 649008,
      "usefulUtilization": 0.637009096960284,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.009762207009112754,
      "meanCycles": 27042.0,
      "p50Cycles": 27042,
      "p95Cycles": 27042,
      "maximumCycles": 27042,
      "firstOffset": 349278,
      "lastOffset": 378048
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c48_r145_r146",
      "dimensions": {},
      "sampleCount": 2080,
      "tileCount": 1440,
      "phaseCount": 37,
      "phaseCycles": 23850,
      "phaseMs": 0.0159,
      "workCycles": 15075840,
      "averageActiveTiles": 632.1106918238994,
      "usefulCycles": 10081152.0,
      "estimatedWorkCycles": 15075840,
      "usefulUtilization": 0.6686958736627611,
      "workLaneOccupancy": 0.9221285563751317,
      "deviceUsefulUtilization": 0.28715340442986054,
      "meanCycles": 7248.0,
      "p50Cycles": 7248,
      "p95Cycles": 7248,
      "maximumCycles": 7248,
      "firstOffset": 172890,
      "lastOffset": 285048
    },
    {
      "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": 57150,
      "lastOffset": 12741342
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 4,
      "phaseCycles": 22464,
      "phaseMs": 0.014976,
      "workCycles": 18294984,
      "averageActiveTiles": 814.4134615384615,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18294984,
      "usefulUtilization": 0.8296674104771012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4590301003344482,
      "meanCycles": 18822.0,
      "p50Cycles": 18822,
      "p95Cycles": 18822,
      "maximumCycles": 18822,
      "firstOffset": 86676,
      "lastOffset": 109140
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 22368,
      "phaseMs": 0.014912,
      "workCycles": 9083340,
      "averageActiveTiles": 406.0863733905579,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 9083340,
      "usefulUtilization": 0.8252541466024612,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.22766607576040307,
      "meanCycles": 18690.0,
      "p50Cycles": 18690,
      "p95Cycles": 18690,
      "maximumCycles": 18690,
      "firstOffset": 86658,
      "lastOffset": 109026
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 15364,
      "tileCount": 1472,
      "phaseCount": 208,
      "phaseCycles": 20850,
      "phaseMs": 0.0139,
      "workCycles": 20358648,
      "averageActiveTiles": 976.4339568345324,
      "usefulCycles": 3096104.0,
      "estimatedWorkCycles": 20358648,
      "usefulUtilization": 0.15207807512561738,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10087920967573767,
      "meanCycles": 1325.0877375683415,
      "p50Cycles": 1074,
      "p95Cycles": 2658,
      "maximumCycles": 2658,
      "firstOffset": 44682,
      "lastOffset": 12947790
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r182_p182_c12_p64",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 1,
      "phaseCycles": 19980,
      "phaseMs": 0.01332,
      "workCycles": 1436400,
      "averageActiveTiles": 71.89189189189189,
      "usefulCycles": 78624.0,
      "estimatedWorkCycles": 1436400,
      "usefulUtilization": 0.05473684210526316,
      "workLaneOccupancy": 0.1875,
      "deviceUsefulUtilization": 0.002673325499412456,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3540,
      "lastOffset": 23520
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r184_p184_c12_p64",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 1,
      "phaseCycles": 19968,
      "phaseMs": 0.013312,
      "workCycles": 718200,
      "averageActiveTiles": 35.96754807692308,
      "usefulCycles": 39744.0,
      "estimatedWorkCycles": 718200,
      "usefulUtilization": 0.05533834586466165,
      "workLaneOccupancy": 0.1875,
      "deviceUsefulUtilization": 0.0013521634615384615,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3522,
      "lastOffset": 23490
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r181_p182_c12_p64",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 1,
      "phaseCycles": 19968,
      "phaseMs": 0.013312,
      "workCycles": 718200,
      "averageActiveTiles": 35.96754807692308,
      "usefulCycles": 39096.0,
      "estimatedWorkCycles": 718200,
      "usefulUtilization": 0.05443609022556391,
      "workLaneOccupancy": 0.18646978021978022,
      "deviceUsefulUtilization": 0.0013301173181438127,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3606,
      "lastOffset": 23574
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 3805,
      "tileCount": 1246,
      "phaseCount": 161,
      "phaseCycles": 18696,
      "phaseMs": 0.012464,
      "workCycles": 792042,
      "averageActiveTiles": 42.36424903722722,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 208.15821287779238,
      "p50Cycles": 150,
      "p95Cycles": 294,
      "maximumCycles": 336,
      "firstOffset": 41286,
      "lastOffset": 12948846
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r66_r67",
      "dimensions": {},
      "sampleCount": 1024,
      "tileCount": 1024,
      "phaseCount": 15,
      "phaseCycles": 18024,
      "phaseMs": 0.012016,
      "workCycles": 14002176,
      "averageActiveTiles": 776.8628495339548,
      "usefulCycles": 6928416.0,
      "estimatedWorkCycles": 14002176,
      "usefulUtilization": 0.4948099495392716,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.26114094251143405,
      "meanCycles": 13674.0,
      "p50Cycles": 13674,
      "p95Cycles": 13674,
      "maximumCycles": 13674,
      "firstOffset": 229752,
      "lastOffset": 247776
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r66_r67",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 384,
      "phaseCount": 11,
      "phaseCycles": 17688,
      "phaseMs": 0.011792,
      "workCycles": 5262336,
      "averageActiveTiles": 297.5088195386703,
      "usefulCycles": 2637522.0,
      "estimatedWorkCycles": 5262336,
      "usefulUtilization": 0.5012074485551664,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.10130002470355731,
      "meanCycles": 13704.0,
      "p50Cycles": 13704,
      "p95Cycles": 13704,
      "maximumCycles": 13704,
      "firstOffset": 230076,
      "lastOffset": 247764
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 8,
      "phaseCycles": 17208,
      "phaseMs": 0.011472,
      "workCycles": 13369860,
      "averageActiveTiles": 776.9560669456067,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 13369860,
      "usefulUtilization": 0.6358899794014298,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3356376205202838,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 12768210,
      "lastOffset": 12785418
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 7,
      "phaseCycles": 16908,
      "phaseMs": 0.011272,
      "workCycles": 10528704,
      "averageActiveTiles": 622.7054648687013,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10528704,
      "usefulUtilization": 0.6282619399310685,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.2657759126114728,
      "meanCycles": 16248.0,
      "p50Cycles": 16248,
      "p95Cycles": 16248,
      "maximumCycles": 16248,
      "firstOffset": 12768264,
      "lastOffset": 12785172
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 36,
      "phaseCycles": 16338,
      "phaseMs": 0.010892,
      "workCycles": 4282080,
      "averageActiveTiles": 262.0932794711715,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.037469329969609395,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 147378,
      "lastOffset": 12847674
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 2496,
      "tileCount": 1472,
      "phaseCount": 85,
      "phaseCycles": 15414,
      "phaseMs": 0.010276,
      "workCycles": 17516592,
      "averageActiveTiles": 1136.407940833009,
      "usefulCycles": 9425760.0,
      "estimatedWorkCycles": 17516592,
      "usefulUtilization": 0.5381046724157302,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4154255589215902,
      "meanCycles": 7017.865384615385,
      "p50Cycles": 11292,
      "p95Cycles": 11880,
      "maximumCycles": 12360,
      "firstOffset": 414990,
      "lastOffset": 12925734
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r145_r146",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 160,
      "phaseCount": 2,
      "phaseCycles": 13356,
      "phaseMs": 0.008904,
      "workCycles": 1157760,
      "averageActiveTiles": 86.68463611859838,
      "usefulCycles": 626400.0,
      "estimatedWorkCycles": 1157760,
      "usefulUtilization": 0.5410447761194029,
      "workLaneOccupancy": 0.75,
      "deviceUsefulUtilization": 0.03186159615609985,
      "meanCycles": 7236.0,
      "p50Cycles": 7236,
      "p95Cycles": 7236,
      "maximumCycles": 7236,
      "firstOffset": 172914,
      "lastOffset": 186270
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 27,
      "phaseCycles": 13056,
      "phaseMs": 0.008704,
      "workCycles": 314400,
      "averageActiveTiles": 24.080882352941178,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0016650682011935208,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 147420,
      "lastOffset": 12846930
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k736_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 22,
      "phaseCycles": 12888,
      "phaseMs": 0.008592,
      "workCycles": 671520,
      "averageActiveTiles": 52.10428305400372,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 671520,
      "usefulUtilization": 0.01954074338813438,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0006916823536555744,
      "meanCycles": 8394.0,
      "p50Cycles": 8394,
      "p95Cycles": 8394,
      "maximumCycles": 8394,
      "firstOffset": 12875316,
      "lastOffset": 12888204
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 13,
      "phaseCycles": 12276,
      "phaseMs": 0.008184,
      "workCycles": 2072928,
      "averageActiveTiles": 168.86021505376345,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.024933769674302633,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 12832968,
      "lastOffset": 12845244
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r182_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 1,
      "phaseCycles": 12060,
      "phaseMs": 0.00804,
      "workCycles": 7383312,
      "averageActiveTiles": 612.2149253731343,
      "usefulCycles": 3773952.0,
      "estimatedWorkCycles": 7383312,
      "usefulUtilization": 0.5111462173073548,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2125892277741726,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6912,
      "lastOffset": 18972
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r184_p184_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 2,
      "phaseCycles": 11988,
      "phaseMs": 0.007992,
      "workCycles": 3691656,
      "averageActiveTiles": 307.94594594594594,
      "usefulCycles": 1907712.0,
      "estimatedWorkCycles": 3691656,
      "usefulUtilization": 0.5167632087063366,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10810810810810811,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6942,
      "lastOffset": 18930
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c32_r145_r146",
      "dimensions": {},
      "sampleCount": 512,
      "tileCount": 512,
      "phaseCount": 11,
      "phaseCycles": 11910,
      "phaseMs": 0.00794,
      "workCycles": 2540544,
      "averageActiveTiles": 213.311838790932,
      "usefulCycles": 1308744.0,
      "estimatedWorkCycles": 2540544,
      "usefulUtilization": 0.5151432134220073,
      "workLaneOccupancy": 0.7294921875,
      "deviceUsefulUtilization": 0.07465091446719965,
      "meanCycles": 4962.0,
      "p50Cycles": 4962,
      "p95Cycles": 4962,
      "maximumCycles": 4962,
      "firstOffset": 170982,
      "lastOffset": 182892
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r181_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 1,
      "phaseCycles": 11868,
      "phaseMs": 0.007912,
      "workCycles": 3759696,
      "averageActiveTiles": 316.7927199191102,
      "usefulCycles": 1876608.0,
      "estimatedWorkCycles": 3759696,
      "usefulUtilization": 0.4991382281971734,
      "workLaneOccupancy": 0.9945054945054945,
      "deviceUsefulUtilization": 0.10742075878137776,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 6870,
      "lastOffset": 18738
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 6,
      "phaseCycles": 10956,
      "phaseMs": 0.007304,
      "workCycles": 146592,
      "averageActiveTiles": 13.380065717415116,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0009921107354318459,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 12831444,
      "lastOffset": 12842400
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 25,
      "phaseCycles": 7818,
      "phaseMs": 0.005212,
      "workCycles": 675648,
      "averageActiveTiles": 86.42210283960092,
      "usefulCycles": 6561.0,
      "estimatedWorkCycles": 675648,
      "usefulUtilization": 0.009710677749360614,
      "workLaneOccupancy": 0.7428668478260869,
      "deviceUsefulUtilization": 0.0005701203743868665,
      "meanCycles": 918.0,
      "p50Cycles": 918,
      "p95Cycles": 918,
      "maximumCycles": 918,
      "firstOffset": 12853260,
      "lastOffset": 12861078
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 35,
      "phaseCycles": 7722,
      "phaseMs": 0.005148,
      "workCycles": 4406112,
      "averageActiveTiles": 570.5920745920746,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2961.0967741935483,
      "p50Cycles": 2976,
      "p95Cycles": 2976,
      "maximumCycles": 2976,
      "firstOffset": 254514,
      "lastOffset": 12890412
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 13,
      "phaseCycles": 6012,
      "phaseMs": 0.004008,
      "workCycles": 1729920,
      "averageActiveTiles": 287.74451097804393,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1729920,
      "usefulUtilization": 0.22641509433962265,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.044259307472012496,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 142278,
      "lastOffset": 148290
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c32_r145_r146",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 2,
      "phaseCycles": 5160,
      "phaseMs": 0.00344,
      "workCycles": 634368,
      "averageActiveTiles": 122.93953488372092,
      "usefulCycles": 324945.0,
      "estimatedWorkCycles": 634368,
      "usefulUtilization": 0.5122342236682809,
      "workLaneOccupancy": 0.7294921875,
      "deviceUsefulUtilization": 0.04278113940849343,
      "meanCycles": 4956.0,
      "p50Cycles": 4956,
      "p95Cycles": 4956,
      "maximumCycles": 4956,
      "firstOffset": 171096,
      "lastOffset": 176256
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 63,
      "phaseCycles": 4776,
      "phaseMs": 0.003184,
      "workCycles": 2324160,
      "averageActiveTiles": 486.63316582914575,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2324160,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.00550988638846406,
      "meanCycles": 1614.0,
      "p50Cycles": 1614,
      "p95Cycles": 1614,
      "maximumCycles": 1614,
      "firstOffset": 12936912,
      "lastOffset": 12943128
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 1,
      "phaseCycles": 4614,
      "phaseMs": 0.003076,
      "workCycles": 4348080,
      "averageActiveTiles": 942.3667100130039,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4348080,
      "usefulUtilization": 0.6633134624937904,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.4246498134222876,
      "meanCycles": 4026.0,
      "p50Cycles": 4026,
      "p95Cycles": 4026,
      "maximumCycles": 4026,
      "firstOffset": 30498,
      "lastOffset": 35112
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 4,
      "phaseCycles": 4566,
      "phaseMs": 0.003044,
      "workCycles": 1458000,
      "averageActiveTiles": 319.31668856767413,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1458000,
      "usefulUtilization": 0.6678518518518518,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.14487516425755584,
      "meanCycles": 4050.0,
      "p50Cycles": 4050,
      "p95Cycles": 4050,
      "maximumCycles": 4050,
      "firstOffset": 30444,
      "lastOffset": 35010
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c224_p224",
      "dimensions": {},
      "sampleCount": 1368,
      "tileCount": 1368,
      "phaseCount": 61,
      "phaseCycles": 4080,
      "phaseMs": 0.00272,
      "workCycles": 2766096,
      "averageActiveTiles": 677.964705882353,
      "usefulCycles": 153216.0,
      "estimatedWorkCycles": 2766096,
      "usefulUtilization": 0.05539070227497527,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.02551150895140665,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 12929130,
      "lastOffset": 12933210
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 12,
      "phaseCycles": 3996,
      "phaseMs": 0.002664,
      "workCycles": 1356960,
      "averageActiveTiles": 339.57957957957956,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1356960,
      "usefulUtilization": 0.3320363164721141,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07659833746790269,
      "meanCycles": 1542.0,
      "p50Cycles": 1542,
      "p95Cycles": 1542,
      "maximumCycles": 1542,
      "firstOffset": 145836,
      "lastOffset": 149832
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 3,
      "phaseCycles": 3138,
      "phaseMs": 0.002092,
      "workCycles": 142464,
      "averageActiveTiles": 45.39961759082218,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 142464,
      "usefulUtilization": 0.19811320754716982,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.00611023360212819,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 142422,
      "lastOffset": 145560
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 15,
      "phaseCycles": 2838,
      "phaseMs": 0.001892,
      "workCycles": 90240,
      "averageActiveTiles": 31.797040169133194,
      "usefulCycles": 1728.0,
      "estimatedWorkCycles": 90240,
      "usefulUtilization": 0.019148936170212766,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0004136409596470264,
      "meanCycles": 2820.0,
      "p50Cycles": 2820,
      "p95Cycles": 2820,
      "maximumCycles": 2820,
      "firstOffset": 12904932,
      "lastOffset": 12907770
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 7,
      "phaseCycles": 2268,
      "phaseMs": 0.001512,
      "workCycles": 142080,
      "averageActiveTiles": 62.645502645502646,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 142080,
      "usefulUtilization": 0.11261261261261261,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.004792577256345372,
      "meanCycles": 1776.0,
      "p50Cycles": 1776,
      "p95Cycles": 1776,
      "maximumCycles": 1776,
      "firstOffset": 147618,
      "lastOffset": 149886
    },
    {
      "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": 12931182,
      "lastOffset": 12933228
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 15,
      "phaseCycles": 2004,
      "phaseMs": 0.001336,
      "workCycles": 39936,
      "averageActiveTiles": 19.92814371257485,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 39936,
      "usefulUtilization": 0.057692307692307696,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0007810466024472793,
      "meanCycles": 1248.0,
      "p50Cycles": 1248,
      "p95Cycles": 1248,
      "maximumCycles": 1248,
      "firstOffset": 12902514,
      "lastOffset": 12904518
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c112_r2_r2",
      "dimensions": {},
      "sampleCount": 1044,
      "tileCount": 1044,
      "phaseCount": 53,
      "phaseCycles": 1752,
      "phaseMs": 0.001168,
      "workCycles": 1722600,
      "averageActiveTiles": 983.2191780821918,
      "usefulCycles": 29232.0,
      "estimatedWorkCycles": 1722600,
      "usefulUtilization": 0.01696969696969697,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.01133487194758785,
      "meanCycles": 1650.0,
      "p50Cycles": 1650,
      "p95Cycles": 1650,
      "maximumCycles": 1650,
      "firstOffset": 12913662,
      "lastOffset": 12915414
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 57,
      "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": 12812724,
      "lastOffset": 12898026
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c96_r2_r2",
      "dimensions": {},
      "sampleCount": 396,
      "tileCount": 396,
      "phaseCount": 43,
      "phaseCycles": 1620,
      "phaseMs": 0.00108,
      "workCycles": 582120,
      "averageActiveTiles": 359.3333333333333,
      "usefulCycles": 9504.0,
      "estimatedWorkCycles": 582120,
      "usefulUtilization": 0.0163265306122449,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.003985507246376811,
      "meanCycles": 1470.0,
      "p50Cycles": 1470,
      "p95Cycles": 1470,
      "maximumCycles": 1470,
      "firstOffset": 12913554,
      "lastOffset": 12915174
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p1_c72_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 5,
      "phaseCycles": 1614,
      "phaseMs": 0.001076,
      "workCycles": 12288,
      "averageActiveTiles": 7.613382899628252,
      "usefulCycles": 576.0,
      "estimatedWorkCycles": 12288,
      "usefulUtilization": 0.046875,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.00024244383384515922,
      "meanCycles": 768.0,
      "p50Cycles": 768,
      "p95Cycles": 768,
      "maximumCycles": 768,
      "firstOffset": 12822828,
      "lastOffset": 12825474
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 43,
      "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": 12892950,
      "lastOffset": 12894360
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c32_p32",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 56,
      "phaseCycles": 1170,
      "phaseMs": 0.00078,
      "workCycles": 838080,
      "averageActiveTiles": 716.3076923076923,
      "usefulCycles": 23040.0,
      "estimatedWorkCycles": 838080,
      "usefulUtilization": 0.027491408934707903,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.013377926421404682,
      "meanCycles": 582.0,
      "p50Cycles": 582,
      "p95Cycles": 582,
      "maximumCycles": 582,
      "firstOffset": 12909456,
      "lastOffset": 12910626
    }
  ],
  "samples": []
}
