{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 241056,
  "matchedSampleCount": 240120,
  "profileSpanCycles": 17198316,
  "profileSpanMs": 11.465544,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 9,
      "phaseCycles": 16260696,
      "phaseMs": 10.840464,
      "workCycles": 23930112390,
      "averageActiveTiles": 1471.6536358591293,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 16256869.830163043,
      "p50Cycles": 16256898,
      "p95Cycles": 16256970,
      "maximumCycles": 16256994,
      "firstOffset": 729660,
      "lastOffset": 16990356
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 16228,
      "tileCount": 1465,
      "phaseCount": 58,
      "phaseCycles": 713484,
      "phaseMs": 0.475656,
      "workCycles": 76200222,
      "averageActiveTiles": 106.80018332576483,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 4695.60155287158,
      "p50Cycles": 2238,
      "p95Cycles": 19704,
      "maximumCycles": 88968,
      "firstOffset": 0,
      "lastOffset": 17194776
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 94172,
      "tileCount": 1472,
      "phaseCount": 75,
      "phaseCycles": 615666,
      "phaseMs": 0.410444,
      "workCycles": 601544898,
      "averageActiveTiles": 977.063696874604,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6387.725629698849,
      "p50Cycles": 4038,
      "p95Cycles": 23430,
      "maximumCycles": 39222,
      "firstOffset": 0,
      "lastOffset": 17195934
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 22066,
      "tileCount": 1472,
      "phaseCount": 145,
      "phaseCycles": 116370,
      "phaseMs": 0.07758,
      "workCycles": 81713442,
      "averageActiveTiles": 702.1864913637536,
      "usefulCycles": 26314664.0,
      "estimatedWorkCycles": 81713442,
      "usefulUtilization": 0.3220359264758422,
      "workLaneOccupancy": 0.9895200907294152,
      "deviceUsefulUtilization": 0.15362043295186642,
      "meanCycles": 3703.1379497870025,
      "p50Cycles": 2154,
      "p95Cycles": 8748,
      "maximumCycles": 10386,
      "firstOffset": 24684,
      "lastOffset": 17193474
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 36202,
      "tileCount": 1472,
      "phaseCount": 184,
      "phaseCycles": 99954,
      "phaseMs": 0.066636,
      "workCycles": 12926658,
      "averageActiveTiles": 129.3260699921964,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 7751250,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 357.07027236064306,
      "p50Cycles": 306,
      "p95Cycles": 744,
      "maximumCycles": 3282,
      "firstOffset": 3048,
      "lastOffset": 17185812
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c192_r72_r76",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 14,
      "phaseCycles": 84000,
      "phaseMs": 0.056,
      "workCycles": 24816240,
      "averageActiveTiles": 295.43142857142857,
      "usefulCycles": 14719680.0,
      "estimatedWorkCycles": 24816240,
      "usefulUtilization": 0.593147068210172,
      "workLaneOccupancy": 0.9340277777777778,
      "deviceUsefulUtilization": 0.11904503105590063,
      "meanCycles": 68934.0,
      "p50Cycles": 68934,
      "p95Cycles": 68934,
      "maximumCycles": 68934,
      "firstOffset": 626916,
      "lastOffset": 710916
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c192_r72_r76",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 38,
      "phaseCycles": 82404,
      "phaseMs": 0.054936,
      "workCycles": 72738000,
      "averageActiveTiles": 882.6998689384011,
      "usefulCycles": 41757408.0,
      "estimatedWorkCycles": 72738000,
      "usefulUtilization": 0.5740796832467211,
      "workLaneOccupancy": 0.9322980967078189,
      "deviceUsefulUtilization": 0.3442527589416301,
      "meanCycles": 67350.0,
      "p50Cycles": 67350,
      "p95Cycles": 67350,
      "maximumCycles": 67350,
      "firstOffset": 626934,
      "lastOffset": 709338
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 9598,
      "tileCount": 1472,
      "phaseCount": 109,
      "phaseCycles": 54468,
      "phaseMs": 0.036312,
      "workCycles": 36810930,
      "averageActiveTiles": 675.8267239480061,
      "usefulCycles": 1972820.0,
      "estimatedWorkCycles": 34682022,
      "usefulUtilization": 0.056883073311008225,
      "workLaneOccupancy": 0.9384120980109366,
      "deviceUsefulUtilization": 0.024605841563135594,
      "meanCycles": 3835.2708897687016,
      "p50Cycles": 3642,
      "p95Cycles": 10950,
      "maximumCycles": 10950,
      "firstOffset": 0,
      "lastOffset": 17184966
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 15,
      "phaseCycles": 53706,
      "phaseMs": 0.035804,
      "workCycles": 71468532,
      "averageActiveTiles": 1330.736454027483,
      "usefulCycles": 33499008.0,
      "estimatedWorkCycles": 71468532,
      "usefulUtilization": 0.46872388535978327,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.4237418214141728,
      "meanCycles": 48551.99184782609,
      "p50Cycles": 48552,
      "p95Cycles": 48552,
      "maximumCycles": 48552,
      "firstOffset": 273306,
      "lastOffset": 327012
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c80_r160_r164",
      "dimensions": {},
      "sampleCount": 795,
      "tileCount": 795,
      "phaseCount": 12,
      "phaseCycles": 50442,
      "phaseMs": 0.033628,
      "workCycles": 39772260,
      "averageActiveTiles": 788.4750802902344,
      "usefulCycles": 31291200.0,
      "estimatedWorkCycles": 39772260,
      "usefulUtilization": 0.7867594147277525,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.42142676134277335,
      "meanCycles": 50028.0,
      "p50Cycles": 50028,
      "p95Cycles": 50028,
      "maximumCycles": 50028,
      "firstOffset": 490140,
      "lastOffset": 540582
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c80_r160_r164",
      "dimensions": {},
      "sampleCount": 636,
      "tileCount": 636,
      "phaseCount": 15,
      "phaseCycles": 49500,
      "phaseMs": 0.033,
      "workCycles": 31233960,
      "averageActiveTiles": 630.9890909090909,
      "usefulCycles": 24346080.0,
      "estimatedWorkCycles": 31233960,
      "usefulUtilization": 0.7794746487477092,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.3341304347826087,
      "meanCycles": 49110.0,
      "p50Cycles": 49110,
      "p95Cycles": 49110,
      "maximumCycles": 49110,
      "firstOffset": 490158,
      "lastOffset": 539658
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 12227,
      "tileCount": 1472,
      "phaseCount": 120,
      "phaseCycles": 48942,
      "phaseMs": 0.032628,
      "workCycles": 16046712,
      "averageActiveTiles": 327.8720117690327,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1312.3997709986097,
      "p50Cycles": 834,
      "p95Cycles": 3018,
      "maximumCycles": 10578,
      "firstOffset": 94758,
      "lastOffset": 17187996
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 1,
      "phaseCycles": 41796,
      "phaseMs": 0.027864,
      "workCycles": 30011040,
      "averageActiveTiles": 718.0361757105943,
      "usefulCycles": 23016960.0,
      "estimatedWorkCycles": 30011040,
      "usefulUtilization": 0.7669497624874047,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3741152679474216,
      "meanCycles": 41682.0,
      "p50Cycles": 41682,
      "p95Cycles": 41682,
      "maximumCycles": 41682,
      "firstOffset": 131964,
      "lastOffset": 173760
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 15,
      "tileCount": 15,
      "phaseCount": 9,
      "phaseCycles": 41772,
      "phaseMs": 0.027848,
      "workCycles": 601470,
      "averageActiveTiles": 14.398879632289573,
      "usefulCycles": 472320.0,
      "estimatedWorkCycles": 601470,
      "usefulUtilization": 0.7852760736196319,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.007681450857449758,
      "meanCycles": 40098.0,
      "p50Cycles": 40098,
      "p95Cycles": 40098,
      "maximumCycles": 40098,
      "firstOffset": 488676,
      "lastOffset": 530448
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 7057,
      "tileCount": 1472,
      "phaseCount": 84,
      "phaseCycles": 41430,
      "phaseMs": 0.02762,
      "workCycles": 9451128,
      "averageActiveTiles": 228.1228095582911,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1339.2557744083888,
      "p50Cycles": 906,
      "p95Cycles": 3060,
      "maximumCycles": 3060,
      "firstOffset": 23778,
      "lastOffset": 17114574
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 12,
      "tileCount": 12,
      "phaseCount": 11,
      "phaseCycles": 41172,
      "phaseMs": 0.027448,
      "workCycles": 472392,
      "averageActiveTiles": 11.47362285048091,
      "usefulCycles": 367488.0,
      "estimatedWorkCycles": 472392,
      "usefulUtilization": 0.7779301935680536,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.006063639704484686,
      "meanCycles": 39366.0,
      "p50Cycles": 39366,
      "p95Cycles": 39366,
      "maximumCycles": 39366,
      "firstOffset": 488664,
      "lastOffset": 529836
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 3,
      "phaseCycles": 41166,
      "phaseMs": 0.027444,
      "workCycles": 29583360,
      "averageActiveTiles": 718.6357673808483,
      "usefulCycles": 22332672.0,
      "estimatedWorkCycles": 29583360,
      "usefulUtilization": 0.7549065420560748,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.3685481264614741,
      "meanCycles": 41088.0,
      "p50Cycles": 41088,
      "p95Cycles": 41088,
      "maximumCycles": 41088,
      "firstOffset": 131976,
      "lastOffset": 173142
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c576_r20_r24",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 22,
      "phaseCycles": 33156,
      "phaseMs": 0.022104,
      "workCycles": 43951680,
      "averageActiveTiles": 1325.6026058631921,
      "usefulCycles": 14929920.0,
      "estimatedWorkCycles": 43951680,
      "usefulUtilization": 0.33968940436406525,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.30590567908228294,
      "meanCycles": 30522.0,
      "p50Cycles": 30522,
      "p95Cycles": 30522,
      "maximumCycles": 30522,
      "firstOffset": 394206,
      "lastOffset": 427362
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c576_r20_r24",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 2,
      "phaseCycles": 30708,
      "phaseMs": 0.020472,
      "workCycles": 678096,
      "averageActiveTiles": 22.0820633059789,
      "usefulCycles": 186624.0,
      "estimatedWorkCycles": 678096,
      "usefulUtilization": 0.27521766829475475,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.004128650797696111,
      "meanCycles": 28254.0,
      "p50Cycles": 28254,
      "p95Cycles": 28254,
      "maximumCycles": 28254,
      "firstOffset": 394218,
      "lastOffset": 424926
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 4,
      "phaseCycles": 29112,
      "phaseMs": 0.019408,
      "workCycles": 20126880,
      "averageActiveTiles": 691.3602638087386,
      "usefulCycles": 15344640.0,
      "estimatedWorkCycles": 20126880,
      "usefulUtilization": 0.7623953638119768,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3580773504426682,
      "meanCycles": 27954.0,
      "p50Cycles": 27954,
      "p95Cycles": 27954,
      "maximumCycles": 27954,
      "firstOffset": 17021670,
      "lastOffset": 17050782
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 7,
      "phaseCycles": 28254,
      "phaseMs": 0.018836,
      "workCycles": 19846080,
      "averageActiveTiles": 702.4166489700573,
      "usefulCycles": 14888448.0,
      "estimatedWorkCycles": 19846080,
      "usefulUtilization": 0.7501959077057031,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.3579824018761484,
      "meanCycles": 27564.0,
      "p50Cycles": 27564,
      "p95Cycles": 27564,
      "maximumCycles": 27564,
      "firstOffset": 17022060,
      "lastOffset": 17050314
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11220,
      "tileCount": 1472,
      "phaseCount": 153,
      "phaseCycles": 26790,
      "phaseMs": 0.01786,
      "workCycles": 20423970,
      "averageActiveTiles": 762.3729003359463,
      "usefulCycles": 4623400.0,
      "estimatedWorkCycles": 20423970,
      "usefulUtilization": 0.22637126866128376,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11724138630572732,
      "meanCycles": 1820.3181818181818,
      "p50Cycles": 1374,
      "p95Cycles": 4314,
      "maximumCycles": 4314,
      "firstOffset": 91530,
      "lastOffset": 17198316
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c112_r121_r122",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 384,
      "phaseCount": 4,
      "phaseCycles": 23874,
      "phaseMs": 0.015916,
      "workCycles": 9050112,
      "averageActiveTiles": 379.07816034179444,
      "usefulCycles": 5902848.0,
      "estimatedWorkCycles": 9050112,
      "usefulUtilization": 0.6522403258655805,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.1679687926833266,
      "meanCycles": 23568.0,
      "p50Cycles": 23568,
      "p95Cycles": 23568,
      "maximumCycles": 23568,
      "firstOffset": 242202,
      "lastOffset": 266076
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c112_r121_r122",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 384,
      "phaseCount": 4,
      "phaseCycles": 23802,
      "phaseMs": 0.015868,
      "workCycles": 9033984,
      "averageActiveTiles": 379.5472649357197,
      "usefulCycles": 5854464.0,
      "estimatedWorkCycles": 9033984,
      "usefulUtilization": 0.6480489671002295,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.16709593274953147,
      "meanCycles": 23526.0,
      "p50Cycles": 23526,
      "p95Cycles": 23526,
      "maximumCycles": 23526,
      "firstOffset": 242220,
      "lastOffset": 266022
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 4374,
      "tileCount": 1458,
      "phaseCount": 34,
      "phaseCycles": 22872,
      "phaseMs": 0.015248,
      "workCycles": 29463264,
      "averageActiveTiles": 1288.180482686254,
      "usefulCycles": 11967264.0,
      "estimatedWorkCycles": 29463264,
      "usefulUtilization": 0.40617577197149646,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3554536018978968,
      "meanCycles": 6736.0,
      "p50Cycles": 6738,
      "p95Cycles": 6738,
      "maximumCycles": 6738,
      "firstOffset": 100014,
      "lastOffset": 17000088
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c96_r121_r122",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 288,
      "phaseCount": 3,
      "phaseCycles": 21060,
      "phaseMs": 0.01404,
      "workCycles": 5833728,
      "averageActiveTiles": 277.0051282051282,
      "usefulCycles": 3702456.0,
      "estimatedWorkCycles": 5833728,
      "usefulUtilization": 0.6346638033175356,
      "workLaneOccupancy": 0.878125,
      "deviceUsefulUtilization": 0.11943283166109253,
      "meanCycles": 20256.0,
      "p50Cycles": 20256,
      "p95Cycles": 20256,
      "maximumCycles": 20256,
      "firstOffset": 241632,
      "lastOffset": 262692
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c96_r121_r122",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 288,
      "phaseCount": 5,
      "phaseCycles": 20502,
      "phaseMs": 0.013668,
      "workCycles": 5823360,
      "averageActiveTiles": 284.03863037752416,
      "usefulCycles": 3672108.0,
      "estimatedWorkCycles": 5823360,
      "usefulUtilization": 0.6305823442136499,
      "workLaneOccupancy": 0.878125,
      "deviceUsefulUtilization": 0.12167781616215598,
      "meanCycles": 20220.0,
      "p50Cycles": 20220,
      "p95Cycles": 20220,
      "maximumCycles": 20220,
      "firstOffset": 241614,
      "lastOffset": 262116
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 30,
      "phaseCycles": 17088,
      "phaseMs": 0.011392,
      "workCycles": 24284448,
      "averageActiveTiles": 1421.1404494382023,
      "usefulCycles": 17256888.0,
      "estimatedWorkCycles": 24284448,
      "usefulUtilization": 0.7106147934678194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.6860621107107963,
      "meanCycles": 16656.0,
      "p50Cycles": 16656,
      "p95Cycles": 16656,
      "maximumCycles": 16656,
      "firstOffset": 567402,
      "lastOffset": 584490
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r104_r105",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 14,
      "phaseCycles": 14370,
      "phaseMs": 0.00958,
      "workCycles": 14086656,
      "averageActiveTiles": 980.282254697286,
      "usefulCycles": 8188128.0,
      "estimatedWorkCycles": 14086656,
      "usefulUtilization": 0.5812684003925417,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3870972133974766,
      "meanCycles": 12228.0,
      "p50Cycles": 12228,
      "p95Cycles": 12228,
      "maximumCycles": 12228,
      "firstOffset": 334716,
      "lastOffset": 349086
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r104_r105",
      "dimensions": {},
      "sampleCount": 192,
      "tileCount": 192,
      "phaseCount": 6,
      "phaseCycles": 14190,
      "phaseMs": 0.00946,
      "workCycles": 2356992,
      "averageActiveTiles": 166.10232558139535,
      "usefulCycles": 1377810.0,
      "estimatedWorkCycles": 2356992,
      "usefulUtilization": 0.584562866568915,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.06596280678371173,
      "meanCycles": 12276.0,
      "p50Cycles": 12276,
      "p95Cycles": 12276,
      "maximumCycles": 12276,
      "firstOffset": 335238,
      "lastOffset": 349428
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c32_p32_b64x16",
      "dimensions": {},
      "sampleCount": 704,
      "tileCount": 704,
      "phaseCount": 6,
      "phaseCycles": 13158,
      "phaseMs": 0.008772,
      "workCycles": 7569408,
      "averageActiveTiles": 575.2704058367533,
      "usefulCycles": 720896.0,
      "estimatedWorkCycles": 7569408,
      "usefulUtilization": 0.09523809523809523,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.03721987615403424,
      "meanCycles": 10752.0,
      "p50Cycles": 10752,
      "p95Cycles": 10752,
      "maximumCycles": 10752,
      "firstOffset": 212826,
      "lastOffset": 225984
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k736_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 160,
      "phaseCount": 14,
      "phaseCycles": 12918,
      "phaseMs": 0.008612,
      "workCycles": 1343040,
      "averageActiveTiles": 103.96655829075708,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1343040,
      "usefulUtilization": 0.01954074338813438,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0013801520628445648,
      "meanCycles": 8394.0,
      "p50Cycles": 8394,
      "p95Cycles": 8394,
      "maximumCycles": 8394,
      "firstOffset": 17129910,
      "lastOffset": 17142828
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c32_p32_b64x16",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 1,
      "phaseCycles": 11778,
      "phaseMs": 0.007852,
      "workCycles": 651264,
      "averageActiveTiles": 55.29495669893021,
      "usefulCycles": 25600.0,
      "estimatedWorkCycles": 651264,
      "usefulUtilization": 0.03930817610062893,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0014765923202433424,
      "meanCycles": 10176.0,
      "p50Cycles": 10176,
      "p95Cycles": 10176,
      "maximumCycles": 10176,
      "firstOffset": 212580,
      "lastOffset": 224358
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 27,
      "phaseCycles": 11580,
      "phaseMs": 0.00772,
      "workCycles": 1854720,
      "averageActiveTiles": 160.1658031088083,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1854720,
      "usefulUtilization": 0.014149844720496895,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0015396204100022529,
      "meanCycles": 1260.0,
      "p50Cycles": 1260,
      "p95Cycles": 1260,
      "maximumCycles": 1260,
      "firstOffset": 17104254,
      "lastOffset": 17115834
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r128_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 800,
      "tileCount": 800,
      "phaseCount": 22,
      "phaseCycles": 10992,
      "phaseMs": 0.007328,
      "workCycles": 3408000,
      "averageActiveTiles": 310.04366812227073,
      "usefulCycles": 819200.0,
      "estimatedWorkCycles": 3408000,
      "usefulUtilization": 0.2403755868544601,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05062970698057085,
      "meanCycles": 4260.0,
      "p50Cycles": 4260,
      "p95Cycles": 4260,
      "maximumCycles": 4260,
      "firstOffset": 17091744,
      "lastOffset": 17102736
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r89_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 160,
      "phaseCount": 11,
      "phaseCycles": 9798,
      "phaseMs": 0.006532,
      "workCycles": 606720,
      "averageActiveTiles": 61.92284139620331,
      "usefulCycles": 113920.0,
      "estimatedWorkCycles": 606720,
      "usefulUtilization": 0.1877637130801688,
      "workLaneOccupancy": 0.6953125,
      "deviceUsefulUtilization": 0.007898683848522768,
      "meanCycles": 3792.0,
      "p50Cycles": 3792,
      "p95Cycles": 3792,
      "maximumCycles": 3792,
      "firstOffset": 17091672,
      "lastOffset": 17101470
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 23,
      "phaseCycles": 9084,
      "phaseMs": 0.006056,
      "workCycles": 6438528,
      "averageActiveTiles": 708.7767503302509,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 4374.0,
      "p50Cycles": 4374,
      "p95Cycles": 4374,
      "maximumCycles": 4374,
      "firstOffset": 358878,
      "lastOffset": 367962
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c48_r240_r244",
      "dimensions": {},
      "sampleCount": 1200,
      "tileCount": 1200,
      "phaseCount": 4,
      "phaseCycles": 8076,
      "phaseMs": 0.005384,
      "workCycles": 8791200,
      "averageActiveTiles": 1088.558692421991,
      "usefulCycles": 6456240.0,
      "estimatedWorkCycles": 8791200,
      "usefulUtilization": 0.7343980343980344,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.5430946766586989,
      "meanCycles": 7326.0,
      "p50Cycles": 7326,
      "p95Cycles": 7326,
      "maximumCycles": 7326,
      "firstOffset": 14760,
      "lastOffset": 22836
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c48_r240_r244",
      "dimensions": {},
      "sampleCount": 240,
      "tileCount": 240,
      "phaseCount": 2,
      "phaseCycles": 7986,
      "phaseMs": 0.005324,
      "workCycles": 1743840,
      "averageActiveTiles": 218.36213373403456,
      "usefulCycles": 1259496.0,
      "estimatedWorkCycles": 1743840,
      "usefulUtilization": 0.7222543352601156,
      "workLaneOccupancy": 0.91109375,
      "deviceUsefulUtilization": 0.10714198216443994,
      "meanCycles": 7266.0,
      "p50Cycles": 7266,
      "p95Cycles": 7266,
      "maximumCycles": 7266,
      "firstOffset": 14772,
      "lastOffset": 22758
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 6138,
      "phaseMs": 0.004092,
      "workCycles": 195072,
      "averageActiveTiles": 31.781036168132943,
      "usefulCycles": 43392.0,
      "estimatedWorkCycles": 195072,
      "usefulUtilization": 0.22244094488188976,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.004802584045220791,
      "meanCycles": 6096.0,
      "p50Cycles": 6096,
      "p95Cycles": 6096,
      "maximumCycles": 6096,
      "firstOffset": 17116998,
      "lastOffset": 17123136
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 356,
      "tileCount": 208,
      "phaseCount": 43,
      "phaseCycles": 5178,
      "phaseMs": 0.003452,
      "workCycles": 95784,
      "averageActiveTiles": 18.498261877172652,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 269.0561797752809,
      "p50Cycles": 288,
      "p95Cycles": 294,
      "maximumCycles": 300,
      "firstOffset": 198834,
      "lastOffset": 17196084
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 4248,
      "phaseMs": 0.002832,
      "workCycles": 54174,
      "averageActiveTiles": 12.752824858757062,
      "usefulCycles": 25824.0,
      "estimatedWorkCycles": 54174,
      "usefulUtilization": 0.47668623324842174,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0041298206828788995,
      "meanCycles": 3385.875,
      "p50Cycles": 3222,
      "p95Cycles": 3462,
      "maximumCycles": 4164,
      "firstOffset": 17176194,
      "lastOffset": 17180442
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 352,
      "tileCount": 352,
      "phaseCount": 3,
      "phaseCycles": 4026,
      "phaseMs": 0.002684,
      "workCycles": 857472,
      "averageActiveTiles": 212.98360655737704,
      "usefulCycles": 180224.0,
      "estimatedWorkCycles": 857472,
      "usefulUtilization": 0.21018062397372742,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.030411023996198622,
      "meanCycles": 2436.0,
      "p50Cycles": 2436,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 213036,
      "lastOffset": 217062
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 3894,
      "phaseMs": 0.002596,
      "workCycles": 62976,
      "averageActiveTiles": 16.172573189522343,
      "usefulCycles": 6400.0,
      "estimatedWorkCycles": 62976,
      "usefulUtilization": 0.1016260162601626,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0011165449632656707,
      "meanCycles": 1968.0,
      "p50Cycles": 1968,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 212016,
      "lastOffset": 215910
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c32_r4_r4",
      "dimensions": {},
      "sampleCount": 184,
      "tileCount": 184,
      "phaseCount": 21,
      "phaseCycles": 3588,
      "phaseMs": 0.002392,
      "workCycles": 464784,
      "averageActiveTiles": 129.53846153846155,
      "usefulCycles": 17216.0,
      "estimatedWorkCycles": 464784,
      "usefulUtilization": 0.03704086199180694,
      "workLaneOccupancy": 0.48731884057971014,
      "deviceUsefulUtilization": 0.0032596577965198,
      "meanCycles": 2526.0,
      "p50Cycles": 2526,
      "p95Cycles": 2526,
      "maximumCycles": 2526,
      "firstOffset": 17186934,
      "lastOffset": 17190522
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c32_r4_r4",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 30,
      "phaseCycles": 3432,
      "phaseMs": 0.002288,
      "workCycles": 1476960,
      "averageActiveTiles": 430.34965034965035,
      "usefulCycles": 48960.0,
      "estimatedWorkCycles": 1476960,
      "usefulUtilization": 0.03314917127071823,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.009691395560960778,
      "meanCycles": 2172.0,
      "p50Cycles": 2172,
      "p95Cycles": 2172,
      "maximumCycles": 2172,
      "firstOffset": 17169540,
      "lastOffset": 17172972
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 4,
      "phaseCycles": 2916,
      "phaseMs": 0.001944,
      "workCycles": 13440,
      "averageActiveTiles": 4.609053497942387,
      "usefulCycles": 2016.0,
      "estimatedWorkCycles": 13440,
      "usefulUtilization": 0.15,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0004696725711218465,
      "meanCycles": 1680.0,
      "p50Cycles": 1680,
      "p95Cycles": 1680,
      "maximumCycles": 1680,
      "firstOffset": 17160708,
      "lastOffset": 17163624
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 26,
      "phaseCycles": 2592,
      "phaseMs": 0.001728,
      "workCycles": 1012176,
      "averageActiveTiles": 390.5,
      "usefulCycles": 28512.0,
      "estimatedWorkCycles": 1012176,
      "usefulUtilization": 0.028169014084507043,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.007472826086956522,
      "meanCycles": 1278.0,
      "p50Cycles": 1278,
      "p95Cycles": 1278,
      "maximumCycles": 1278,
      "firstOffset": 17169054,
      "lastOffset": 17171646
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 1288,
      "tileCount": 1288,
      "phaseCount": 39,
      "phaseCycles": 2508,
      "phaseMs": 0.001672,
      "workCycles": 1877904,
      "averageActiveTiles": 748.7655502392345,
      "usefulCycles": 60256.0,
      "estimatedWorkCycles": 1877904,
      "usefulUtilization": 0.032086837239816306,
      "workLaneOccupancy": 0.48731884057971014,
      "deviceUsefulUtilization": 0.016321683655779767,
      "meanCycles": 1458.0,
      "p50Cycles": 1458,
      "p95Cycles": 1458,
      "maximumCycles": 1458,
      "firstOffset": 17186610,
      "lastOffset": 17189118
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 40,
      "phaseCycles": 2364,
      "phaseMs": 0.001576,
      "workCycles": 1944000,
      "averageActiveTiles": 822.3350253807107,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1944000,
      "usefulUtilization": 0.021333333333333333,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.011917898918561024,
      "meanCycles": 750.0,
      "p50Cycles": 750,
      "p95Cycles": 750,
      "maximumCycles": 750,
      "firstOffset": 17080128,
      "lastOffset": 17155596
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 4,
      "phaseCycles": 2202,
      "phaseMs": 0.001468,
      "workCycles": 17568,
      "averageActiveTiles": 7.978201634877384,
      "usefulCycles": 3456.0,
      "estimatedWorkCycles": 17568,
      "usefulUtilization": 0.19672131147540983,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0010662243809975121,
      "meanCycles": 2196.0,
      "p50Cycles": 2196,
      "p95Cycles": 2196,
      "maximumCycles": 2196,
      "firstOffset": 17163900,
      "lastOffset": 17166102
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 1638,
      "phaseMs": 0.001092,
      "workCycles": 51072,
      "averageActiveTiles": 31.17948717948718,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1596.0,
      "p50Cycles": 1596,
      "p95Cycles": 1596,
      "maximumCycles": 1596,
      "firstOffset": 17143440,
      "lastOffset": 17145078
    }
  ],
  "samples": []
}
