{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 189838,
  "matchedSampleCount": 97775,
  "profileSpanCycles": 10371582,
  "profileSpanMs": 6.914388,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 9810954,
      "phaseMs": 6.540636,
      "workCycles": 14437312488,
      "averageActiveTiles": 1471.5503189598076,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 9807956.85326087,
      "p50Cycles": 9807960,
      "p95Cycles": 9807972,
      "maximumCycles": 9807990,
      "firstOffset": 414930,
      "lastOffset": 10225884
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 7912,
      "tileCount": 1472,
      "phaseCount": 148,
      "phaseCycles": 62094,
      "phaseMs": 0.041396,
      "workCycles": 7582626,
      "averageActiveTiles": 122.11527683834187,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 958.3703235591506,
      "p50Cycles": 1008,
      "p95Cycles": 1632,
      "maximumCycles": 1686,
      "firstOffset": 27102,
      "lastOffset": 10365072
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10600,
      "tileCount": 1472,
      "phaseCount": 134,
      "phaseCycles": 61926,
      "phaseMs": 0.041284,
      "workCycles": 42208656,
      "averageActiveTiles": 681.5982947388819,
      "usefulCycles": 23067536.0,
      "estimatedWorkCycles": 42208656,
      "usefulUtilization": 0.5465119761216751,
      "workLaneOccupancy": 0.9933504298002166,
      "deviceUsefulUtilization": 0.2530581732193684,
      "meanCycles": 3981.948679245283,
      "p50Cycles": 3546,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 19698,
      "lastOffset": 10367880
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 26518,
      "tileCount": 1472,
      "phaseCount": 152,
      "phaseCycles": 47376,
      "phaseMs": 0.031584,
      "workCycles": 8300508,
      "averageActiveTiles": 175.2049138804458,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 5343984,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 313.0141036277246,
      "p50Cycles": 300,
      "p95Cycles": 564,
      "maximumCycles": 1260,
      "firstOffset": 6738,
      "lastOffset": 10363398
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 5993,
      "tileCount": 1440,
      "phaseCount": 122,
      "phaseCycles": 35268,
      "phaseMs": 0.023512,
      "workCycles": 18627384,
      "averageActiveTiles": 528.1667233752977,
      "usefulCycles": 682842.0,
      "estimatedWorkCycles": 18294108,
      "usefulUtilization": 0.03732578817179826,
      "workLaneOccupancy": 0.9062269409422694,
      "deviceUsefulUtilization": 0.013153204838972143,
      "meanCycles": 3108.19022192558,
      "p50Cycles": 2088,
      "p95Cycles": 9300,
      "maximumCycles": 9300,
      "firstOffset": 0,
      "lastOffset": 10363380
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 30,
      "phaseCycles": 32730,
      "phaseMs": 0.02182,
      "workCycles": 23016240,
      "averageActiveTiles": 703.2153987167736,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23016240,
      "usefulUtilization": 0.6669377795851973,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.31861475311839954,
      "meanCycles": 29508.0,
      "p50Cycles": 29508,
      "p95Cycles": 29508,
      "maximumCycles": 29508,
      "firstOffset": 267276,
      "lastOffset": 300006
    },
    {
      "name": "layer_norm_f8",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 34,
      "phaseCycles": 32460,
      "phaseMs": 0.02164,
      "workCycles": 20855232,
      "averageActiveTiles": 642.4902033271719,
      "usefulCycles": 8503056.0,
      "estimatedWorkCycles": 20855232,
      "usefulUtilization": 0.4077181208053691,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.17795849071767258,
      "meanCycles": 9536.0,
      "p50Cycles": 9540,
      "p95Cycles": 9540,
      "maximumCycles": 9540,
      "firstOffset": 35532,
      "lastOffset": 10241136
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 22,
      "phaseCycles": 30048,
      "phaseMs": 0.020032,
      "workCycles": 18094752,
      "averageActiveTiles": 602.1948881789137,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18094752,
      "usefulUtilization": 0.660045520380716,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.2700244825670232,
      "meanCycles": 28998.0,
      "p50Cycles": 28998,
      "p95Cycles": 28998,
      "maximumCycles": 28998,
      "firstOffset": 269484,
      "lastOffset": 299532
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 17,
      "phaseCycles": 29796,
      "phaseMs": 0.019864,
      "workCycles": 797940,
      "averageActiveTiles": 26.780104712041886,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 797940,
      "usefulUtilization": 0.6659147304308595,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.01211499063194943,
      "meanCycles": 26598.0,
      "p50Cycles": 26598,
      "p95Cycles": 26598,
      "maximumCycles": 26598,
      "firstOffset": 267288,
      "lastOffset": 297084
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 19,
      "phaseCycles": 28590,
      "phaseMs": 0.01906,
      "workCycles": 38108064,
      "averageActiveTiles": 1332.9158447009445,
      "usefulCycles": 16866144.0,
      "estimatedWorkCycles": 38108064,
      "usefulUtilization": 0.4425872697180313,
      "workLaneOccupancy": 0.9972413793103448,
      "deviceUsefulUtilization": 0.4007687394497924,
      "meanCycles": 25888.630434782608,
      "p50Cycles": 25890,
      "p95Cycles": 25890,
      "maximumCycles": 25890,
      "firstOffset": 148356,
      "lastOffset": 176946
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 34,
      "phaseCycles": 28428,
      "phaseMs": 0.018952,
      "workCycles": 31871880,
      "averageActiveTiles": 1121.1439425918109,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 31871880,
      "usefulUtilization": 0.7413732732427456,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5646645071301412,
      "meanCycles": 26232.0,
      "p50Cycles": 26232,
      "p95Cycles": 26232,
      "maximumCycles": 26232,
      "firstOffset": 360432,
      "lastOffset": 388860
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 14,
      "phaseCycles": 27180,
      "phaseMs": 0.01812,
      "workCycles": 627408,
      "averageActiveTiles": 23.08344370860927,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 627408,
      "usefulUtilization": 0.6589396373651595,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.010333285344082925,
      "meanCycles": 26142.0,
      "p50Cycles": 26142,
      "p95Cycles": 26142,
      "maximumCycles": 26142,
      "firstOffset": 269484,
      "lastOffset": 296664
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 21,
      "phaseCycles": 26442,
      "phaseMs": 0.017628,
      "workCycles": 6349590,
      "averageActiveTiles": 240.13274336283186,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6349590,
      "usefulUtilization": 0.7259656135278025,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.11842942551869062,
      "meanCycles": 26130.0,
      "p50Cycles": 26130,
      "p95Cycles": 26130,
      "maximumCycles": 26130,
      "firstOffset": 361998,
      "lastOffset": 388440
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3961,
      "tileCount": 1472,
      "phaseCount": 104,
      "phaseCycles": 23028,
      "phaseMs": 0.015352,
      "workCycles": 1632456,
      "averageActiveTiles": 70.89004689942678,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 412.1322898258016,
      "p50Cycles": 384,
      "p95Cycles": 522,
      "maximumCycles": 804,
      "firstOffset": 49230,
      "lastOffset": 10315296
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 3,
      "phaseCycles": 19008,
      "phaseMs": 0.012672,
      "workCycles": 18055872,
      "averageActiveTiles": 949.9090909090909,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18055872,
      "usefulUtilization": 0.8406546080964685,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5424901185770751,
      "meanCycles": 18576.0,
      "p50Cycles": 18576,
      "p95Cycles": 18576,
      "maximumCycles": 18576,
      "firstOffset": 63546,
      "lastOffset": 82554
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 18720,
      "phaseMs": 0.01248,
      "workCycles": 8963784,
      "averageActiveTiles": 478.8346153846154,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 8963784,
      "usefulUtilization": 0.836261114725656,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.27203177257525085,
      "meanCycles": 18444.0,
      "p50Cycles": 18444,
      "p95Cycles": 18444,
      "maximumCycles": 18444,
      "firstOffset": 63468,
      "lastOffset": 82188
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11800,
      "tileCount": 1458,
      "phaseCount": 148,
      "phaseCycles": 18252,
      "phaseMs": 0.012168,
      "workCycles": 12935772,
      "averageActiveTiles": 708.7317554240631,
      "usefulCycles": 2310624.0,
      "estimatedWorkCycles": 12935772,
      "usefulUtilization": 0.17862281431676438,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.08600248692221936,
      "meanCycles": 1096.2518644067798,
      "p50Cycles": 942,
      "p95Cycles": 1806,
      "maximumCycles": 1806,
      "firstOffset": 24384,
      "lastOffset": 10371582
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 7,
      "phaseCycles": 16614,
      "phaseMs": 0.011076,
      "workCycles": 12956760,
      "averageActiveTiles": 779.8699891657639,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 12956760,
      "usefulUtilization": 0.6561640410102526,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3476376654576287,
      "meanCycles": 15996.0,
      "p50Cycles": 15996,
      "p95Cycles": 15996,
      "maximumCycles": 15996,
      "firstOffset": 10251210,
      "lastOffset": 10267824
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 7,
      "phaseCycles": 16044,
      "phaseMs": 0.010696,
      "workCycles": 10198224,
      "averageActiveTiles": 635.6409872849663,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10198224,
      "usefulUtilization": 0.6486211716863642,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.28008845240805175,
      "meanCycles": 15738.0,
      "p50Cycles": 15738,
      "p95Cycles": 15738,
      "maximumCycles": 15738,
      "firstOffset": 10251384,
      "lastOffset": 10267428
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 11,
      "phaseCycles": 12684,
      "phaseMs": 0.008456,
      "workCycles": 10715328,
      "averageActiveTiles": 844.7909176915799,
      "usefulCycles": 6298560.0,
      "estimatedWorkCycles": 10715328,
      "usefulUtilization": 0.5878084179970973,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.33734729134959485,
      "meanCycles": 12402.0,
      "p50Cycles": 12402,
      "p95Cycles": 12402,
      "maximumCycles": 12402,
      "firstOffset": 129606,
      "lastOffset": 142290
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 12,
      "phaseCycles": 10584,
      "phaseMs": 0.007056,
      "workCycles": 5761152,
      "averageActiveTiles": 544.3265306122449,
      "usefulCycles": 3267378.0,
      "estimatedWorkCycles": 5761152,
      "usefulUtilization": 0.5671396970605879,
      "workLaneOccupancy": 0.875390625,
      "deviceUsefulUtilization": 0.2097209128216504,
      "meanCycles": 10002.0,
      "p50Cycles": 10002,
      "p95Cycles": 10002,
      "maximumCycles": 10002,
      "firstOffset": 129018,
      "lastOffset": 139602
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 950,
      "phaseCount": 29,
      "phaseCycles": 10470,
      "phaseMs": 0.00698,
      "workCycles": 4282080,
      "averageActiveTiles": 408.9856733524355,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0584693326689091,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 114774,
      "lastOffset": 10306470
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 2534,
      "tileCount": 1458,
      "phaseCount": 72,
      "phaseCycles": 9858,
      "phaseMs": 0.006572,
      "workCycles": 13055688,
      "averageActiveTiles": 1324.374923919659,
      "usefulCycles": 8640280.0,
      "estimatedWorkCycles": 13055688,
      "usefulUtilization": 0.6618019670813212,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5954306588336994,
      "meanCycles": 5152.205209155485,
      "p50Cycles": 8556,
      "p95Cycles": 8556,
      "maximumCycles": 8556,
      "firstOffset": 324900,
      "lastOffset": 10360566
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 20,
      "phaseCycles": 9654,
      "phaseMs": 0.006436,
      "workCycles": 9303552,
      "averageActiveTiles": 963.6991920447483,
      "usefulCycles": 4408992.0,
      "estimatedWorkCycles": 9303552,
      "usefulUtilization": 0.4739041604754829,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3102588699435242,
      "meanCycles": 8076.0,
      "p50Cycles": 8076,
      "p95Cycles": 8076,
      "maximumCycles": 8076,
      "firstOffset": 184140,
      "lastOffset": 193794
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 150,
      "phaseCount": 21,
      "phaseCycles": 9222,
      "phaseMs": 0.006148,
      "workCycles": 314400,
      "averageActiveTiles": 34.09238776837996,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0023573119100826945,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 113892,
      "lastOffset": 10305072
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 96,
      "phaseCount": 6,
      "phaseCycles": 8658,
      "phaseMs": 0.005772,
      "workCycles": 779904,
      "averageActiveTiles": 90.07900207900208,
      "usefulCycles": 373977.0,
      "estimatedWorkCycles": 779904,
      "usefulUtilization": 0.47951670974889216,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.02934401269999096,
      "meanCycles": 8124.0,
      "p50Cycles": 8124,
      "p95Cycles": 8124,
      "maximumCycles": 8124,
      "firstOffset": 185190,
      "lastOffset": 193848
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1694,
      "tileCount": 794,
      "phaseCount": 77,
      "phaseCycles": 8208,
      "phaseMs": 0.005472,
      "workCycles": 488592,
      "averageActiveTiles": 59.526315789473685,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 288.4250295159386,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 300,
      "firstOffset": 24378,
      "lastOffset": 10370214
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 18,
      "phaseCycles": 8136,
      "phaseMs": 0.005424,
      "workCycles": 896448,
      "averageActiveTiles": 110.18289085545723,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 896448,
      "usefulUtilization": 0.01463777040051403,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0010956738168526356,
      "meanCycles": 1218.0,
      "p50Cycles": 1218,
      "p95Cycles": 1218,
      "maximumCycles": 1218,
      "firstOffset": 10308378,
      "lastOffset": 10316514
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 29,
      "phaseCycles": 7404,
      "phaseMs": 0.004936,
      "workCycles": 10255680,
      "averageActiveTiles": 1385.1539708265802,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10255680,
      "usefulUtilization": 0.7369839932603202,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.6935029243886971,
      "meanCycles": 7122.0,
      "p50Cycles": 7122,
      "p95Cycles": 7122,
      "maximumCycles": 7122,
      "firstOffset": 218868,
      "lastOffset": 226272
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k112_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 560,
      "tileCount": 560,
      "phaseCount": 22,
      "phaseCycles": 7284,
      "phaseMs": 0.004856,
      "workCycles": 863520,
      "averageActiveTiles": 118.5502471169687,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 863520,
      "usefulUtilization": 0.015195942190105614,
      "workLaneOccupancy": 0.8368622448979591,
      "deviceUsefulUtilization": 0.001223833357209369,
      "meanCycles": 1542.0,
      "p50Cycles": 1542,
      "p95Cycles": 1542,
      "maximumCycles": 1542,
      "firstOffset": 10325112,
      "lastOffset": 10332396
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 32,
      "phaseCycles": 6978,
      "phaseMs": 0.004652,
      "workCycles": 4398336,
      "averageActiveTiles": 630.3147033533963,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2988.0,
      "p50Cycles": 2988,
      "p95Cycles": 2988,
      "maximumCycles": 2988,
      "firstOffset": 198780,
      "lastOffset": 205758
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 6120,
      "phaseMs": 0.00408,
      "workCycles": 97536,
      "averageActiveTiles": 15.937254901960785,
      "usefulCycles": 21696.0,
      "estimatedWorkCycles": 97536,
      "usefulUtilization": 0.22244094488188976,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.0024083546462063087,
      "meanCycles": 6096.0,
      "p50Cycles": 6096,
      "p95Cycles": 6096,
      "maximumCycles": 6096,
      "firstOffset": 10317672,
      "lastOffset": 10323792
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 3,
      "phaseCycles": 4362,
      "phaseMs": 0.002908,
      "workCycles": 1442880,
      "averageActiveTiles": 330.7840440165062,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1442880,
      "usefulUtilization": 0.6748502994011976,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.1516506189821183,
      "meanCycles": 4008.0,
      "p50Cycles": 4008,
      "p95Cycles": 4008,
      "maximumCycles": 4008,
      "firstOffset": 12372,
      "lastOffset": 16734
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 2,
      "phaseCycles": 4308,
      "phaseMs": 0.002872,
      "workCycles": 4302720,
      "averageActiveTiles": 998.7743732590529,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4302720,
      "usefulUtilization": 0.6703062248995983,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.4548129617294417,
      "meanCycles": 3984.0,
      "p50Cycles": 3984,
      "p95Cycles": 3984,
      "maximumCycles": 3984,
      "firstOffset": 12378,
      "lastOffset": 16686
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c32_r2_r2",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 37,
      "phaseCycles": 3306,
      "phaseMs": 0.002204,
      "workCycles": 1391280,
      "averageActiveTiles": 420.8348457350272,
      "usefulCycles": 24480.0,
      "estimatedWorkCycles": 1391280,
      "usefulUtilization": 0.017595307917888565,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.005030379547068571,
      "meanCycles": 2046.0,
      "p50Cycles": 2046,
      "p95Cycles": 2046,
      "maximumCycles": 2046,
      "firstOffset": 10354992,
      "lastOffset": 10358298
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 24,
      "phaseCycles": 2538,
      "phaseMs": 0.001692,
      "workCycles": 969408,
      "averageActiveTiles": 381.9574468085106,
      "usefulCycles": 14256.0,
      "estimatedWorkCycles": 969408,
      "usefulUtilization": 0.014705882352941176,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.0038159111933395006,
      "meanCycles": 1224.0,
      "p50Cycles": 1224,
      "p95Cycles": 1224,
      "maximumCycles": 1224,
      "firstOffset": 10354506,
      "lastOffset": 10357044
    },
    {
      "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": 10346766,
      "lastOffset": 10349106
    },
    {
      "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": 10349382,
      "lastOffset": 10351578
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 38,
      "phaseCycles": 2160,
      "phaseMs": 0.00144,
      "workCycles": 2229120,
      "averageActiveTiles": 1032.0,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2229120,
      "usefulUtilization": 0.017377260981912144,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.012182971014492753,
      "meanCycles": 1548.0,
      "p50Cycles": 1548,
      "p95Cycles": 1548,
      "maximumCycles": 1548,
      "firstOffset": 10364460,
      "lastOffset": 10366620
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 46,
      "phaseCycles": 1680,
      "phaseMs": 0.00112,
      "workCycles": 1912896,
      "averageActiveTiles": 1138.6285714285714,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 1912896,
      "usefulUtilization": 0.01084010840108401,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.008385093167701863,
      "meanCycles": 738.0,
      "p50Cycles": 738,
      "p95Cycles": 738,
      "maximumCycles": 738,
      "firstOffset": 10289490,
      "lastOffset": 10342464
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 1614,
      "phaseMs": 0.001076,
      "workCycles": 25536,
      "averageActiveTiles": 15.821561338289962,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1596.0,
      "p50Cycles": 1596,
      "p95Cycles": 1596,
      "maximumCycles": 1596,
      "firstOffset": 10335174,
      "lastOffset": 10336788
    }
  ],
  "samples": []
}
