{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 187988,
  "matchedSampleCount": 97253,
  "profileSpanCycles": 10158156,
  "profileSpanMs": 6.772104,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 9605028,
      "phaseMs": 6.403352,
      "workCycles": 14134189416,
      "averageActiveTiles": 1471.5406780698609,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 9602030.85326087,
      "p50Cycles": 9602034,
      "p95Cycles": 9602046,
      "maximumCycles": 9602064,
      "firstOffset": 406662,
      "lastOffset": 10011690
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 7917,
      "tileCount": 1472,
      "phaseCount": 148,
      "phaseCycles": 59298,
      "phaseMs": 0.039532,
      "workCycles": 7581906,
      "averageActiveTiles": 127.86107457249823,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 957.6741189844638,
      "p50Cycles": 1008,
      "p95Cycles": 1632,
      "maximumCycles": 1686,
      "firstOffset": 26568,
      "lastOffset": 10151646
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 8958,
      "tileCount": 1472,
      "phaseCount": 127,
      "phaseCycles": 44304,
      "phaseMs": 0.029536,
      "workCycles": 33358212,
      "averageActiveTiles": 752.9390574214518,
      "usefulCycles": 19137984.0,
      "estimatedWorkCycles": 33358212,
      "usefulUtilization": 0.5737113248156106,
      "workLaneOccupancy": 0.9926603860019519,
      "deviceUsefulUtilization": 0.2934576522681238,
      "meanCycles": 3723.8459477561955,
      "p50Cycles": 3048,
      "p95Cycles": 6444,
      "maximumCycles": 6894,
      "firstOffset": 19164,
      "lastOffset": 10154454
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 26522,
      "tileCount": 1472,
      "phaseCount": 157,
      "phaseCycles": 44112,
      "phaseMs": 0.029408,
      "workCycles": 8249058,
      "averageActiveTiles": 187.00258433079435,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 5342808,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 311.0269964557726,
      "p50Cycles": 300,
      "p95Cycles": 564,
      "maximumCycles": 1260,
      "firstOffset": 23052,
      "lastOffset": 10149972
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 29,
      "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": 265722,
      "lastOffset": 298452
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 7997,
      "tileCount": 1458,
      "phaseCount": 130,
      "phaseCycles": 30072,
      "phaseMs": 0.020048,
      "workCycles": 20899932,
      "averageActiveTiles": 694.9964086193137,
      "usefulCycles": 792522.0,
      "estimatedWorkCycles": 20605236,
      "usefulUtilization": 0.03846216563595777,
      "workLaneOccupancy": 0.9148819748434643,
      "deviceUsefulUtilization": 0.017903634537978417,
      "meanCycles": 2613.471551831937,
      "p50Cycles": 2082,
      "p95Cycles": 5496,
      "maximumCycles": 5496,
      "firstOffset": 0,
      "lastOffset": 10149954
    },
    {
      "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": 267930,
      "lastOffset": 297978
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 15,
      "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": 265734,
      "lastOffset": 295530
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 20,
      "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": 148254,
      "lastOffset": 176844
    },
    {
      "name": "reduce_bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1642,
      "tileCount": 1458,
      "phaseCount": 61,
      "phaseCycles": 28176,
      "phaseMs": 0.018784,
      "workCycles": 21773532,
      "averageActiveTiles": 772.7687393526405,
      "usefulCycles": 13355312.0,
      "estimatedWorkCycles": 21773532,
      "usefulUtilization": 0.6133737052858489,
      "workLaneOccupancy": 0.997107969151671,
      "deviceUsefulUtilization": 0.3220081691479644,
      "meanCycles": 13260.372716199756,
      "p50Cycles": 17586,
      "p95Cycles": 17898,
      "maximumCycles": 17898,
      "firstOffset": 304998,
      "lastOffset": 10146576
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 33,
      "phaseCycles": 27582,
      "phaseMs": 0.018388,
      "workCycles": 31871880,
      "averageActiveTiles": 1155.531868609963,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 31871880,
      "usefulUtilization": 0.7413732732427456,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5819839971247789,
      "meanCycles": 26232.0,
      "p50Cycles": 26232,
      "p95Cycles": 26232,
      "maximumCycles": 26232,
      "firstOffset": 354600,
      "lastOffset": 382182
    },
    {
      "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": 267930,
      "lastOffset": 295110
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 19,
      "phaseCycles": 26532,
      "phaseMs": 0.017688,
      "workCycles": 6349590,
      "averageActiveTiles": 239.3181818181818,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6349590,
      "usefulUtilization": 0.7259656135278025,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.11802769748097458,
      "meanCycles": 26130.0,
      "p50Cycles": 26130,
      "p95Cycles": 26130,
      "maximumCycles": 26130,
      "firstOffset": 355578,
      "lastOffset": 382110
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3960,
      "tileCount": 1472,
      "phaseCount": 103,
      "phaseCycles": 21930,
      "phaseMs": 0.01462,
      "workCycles": 1631652,
      "averageActiveTiles": 74.40273597811218,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 412.03333333333336,
      "p50Cycles": 384,
      "p95Cycles": 522,
      "maximumCycles": 720,
      "firstOffset": 103494,
      "lastOffset": 10101102
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 3,
      "phaseCycles": 21516,
      "phaseMs": 0.014344,
      "workCycles": 18055872,
      "averageActiveTiles": 839.1834913552705,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18055872,
      "usefulUtilization": 0.8406546080964685,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.47925507408036083,
      "meanCycles": 18576.0,
      "p50Cycles": 18576,
      "p95Cycles": 18576,
      "maximumCycles": 18576,
      "firstOffset": 60948,
      "lastOffset": 82464
    },
    {
      "name": "layer_norm_f8",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 729,
      "phaseCount": 30,
      "phaseCycles": 20406,
      "phaseMs": 0.013604,
      "workCycles": 13900572,
      "averageActiveTiles": 681.2002352249339,
      "usefulCycles": 5668704.0,
      "estimatedWorkCycles": 13900572,
      "usefulUtilization": 0.407803650094399,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.18872006954476306,
      "meanCycles": 9534.0,
      "p50Cycles": 9528,
      "p95Cycles": 9540,
      "maximumCycles": 9540,
      "firstOffset": 241296,
      "lastOffset": 10026942
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 19614,
      "phaseMs": 0.013076,
      "workCycles": 8963784,
      "averageActiveTiles": 457.009483022331,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 8963784,
      "usefulUtilization": 0.836261114725656,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.25963264926117546,
      "meanCycles": 18444.0,
      "p50Cycles": 18444,
      "p95Cycles": 18444,
      "maximumCycles": 18444,
      "firstOffset": 62718,
      "lastOffset": 82332
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11800,
      "tileCount": 1458,
      "phaseCount": 152,
      "phaseCycles": 18246,
      "phaseMs": 0.012164,
      "workCycles": 12935772,
      "averageActiveTiles": 708.9648142058534,
      "usefulCycles": 2310624.0,
      "estimatedWorkCycles": 12935772,
      "usefulUtilization": 0.17862281431676438,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.08603076791101326,
      "meanCycles": 1096.2518644067798,
      "p50Cycles": 942,
      "p95Cycles": 1806,
      "maximumCycles": 1806,
      "firstOffset": 23856,
      "lastOffset": 10158156
    },
    {
      "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": 10037016,
      "lastOffset": 10053630
    },
    {
      "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": 10037190,
      "lastOffset": 10053234
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 12,
      "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": 129504,
      "lastOffset": 142188
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 13,
      "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": 128916,
      "lastOffset": 139500
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 950,
      "phaseCount": 31,
      "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": 114672,
      "lastOffset": 10092276
    },
    {
      "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": 184038,
      "lastOffset": 193692
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 31,
      "phaseCycles": 9516,
      "phaseMs": 0.006344,
      "workCycles": 10255680,
      "averageActiveTiles": 1077.7301387137452,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10255680,
      "usefulUtilization": 0.7369839932603202,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.5395855035912056,
      "meanCycles": 7122.0,
      "p50Cycles": 7122,
      "p95Cycles": 7122,
      "maximumCycles": 7122,
      "firstOffset": 215172,
      "lastOffset": 224688
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 729,
      "tileCount": 729,
      "phaseCount": 2,
      "phaseCycles": 9252,
      "phaseMs": 0.006168,
      "workCycles": 4912002,
      "averageActiveTiles": 530.9124513618677,
      "usefulCycles": 1994544.0,
      "estimatedWorkCycles": 4912002,
      "usefulUtilization": 0.40605520926090827,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.14645364574522077,
      "meanCycles": 6738.0,
      "p50Cycles": 6738,
      "p95Cycles": 6738,
      "maximumCycles": 6738,
      "firstOffset": 34998,
      "lastOffset": 44250
    },
    {
      "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": 113790,
      "lastOffset": 10090878
    },
    {
      "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": 185088,
      "lastOffset": 193746
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1694,
      "tileCount": 794,
      "phaseCount": 79,
      "phaseCycles": 8226,
      "phaseMs": 0.005484,
      "workCycles": 488592,
      "averageActiveTiles": 59.39606126914661,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 288.4250295159386,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 300,
      "firstOffset": 23850,
      "lastOffset": 10156788
    },
    {
      "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": 10094184,
      "lastOffset": 10102320
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 2,
      "phaseCycles": 7452,
      "phaseMs": 0.004968,
      "workCycles": 4302720,
      "averageActiveTiles": 577.3913043478261,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4302720,
      "usefulUtilization": 0.6703062248995983,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.26292729993698805,
      "meanCycles": 3984.0,
      "p50Cycles": 3984,
      "p95Cycles": 3984,
      "maximumCycles": 3984,
      "firstOffset": 8724,
      "lastOffset": 16176
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 3,
      "phaseCycles": 7404,
      "phaseMs": 0.004936,
      "workCycles": 1442880,
      "averageActiveTiles": 194.87844408427875,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1442880,
      "usefulUtilization": 0.6748502994011976,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.08934359805510535,
      "meanCycles": 4008.0,
      "p50Cycles": 4008,
      "p95Cycles": 4008,
      "maximumCycles": 4008,
      "firstOffset": 8748,
      "lastOffset": 16152
    },
    {
      "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": 10110918,
      "lastOffset": 10118202
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 33,
      "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": 198678,
      "lastOffset": 205656
    },
    {
      "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": 10103478,
      "lastOffset": 10109598
    },
    {
      "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": 10140798,
      "lastOffset": 10144104
    },
    {
      "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": 10140312,
      "lastOffset": 10142850
    },
    {
      "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": 10132572,
      "lastOffset": 10134912
    },
    {
      "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": 10135188,
      "lastOffset": 10137384
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 39,
      "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": 10151034,
      "lastOffset": 10153194
    },
    {
      "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": 10075296,
      "lastOffset": 10128270
    },
    {
      "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": 10120980,
      "lastOffset": 10122594
    }
  ],
  "samples": [
    {
      "physicalTile": 109,
      "offset": 408630,
      "duration": 9602064,
      "phase": 12,
      "epoch": 0,
      "kind": "compute",
      "operation": "operation.24",
      "kernel": "repeat-remainder",
      "metadata": [
        {
          "name": "reason",
          "value": "Repeat"
        },
        {
          "name": "value",
          "value": "434"
        },
        {
          "name": "iterations",
          "value": "26"
        }
      ]
    },
    {
      "physicalTile": 173,
      "offset": 408054,
      "duration": 9602064,
      "phase": 11,
      "epoch": 0,
      "kind": "compute",
      "operation": "operation.24",
      "kernel": "repeat-remainder",
      "metadata": [
        {
          "name": "reason",
          "value": "Repeat"
        },
        {
          "name": "value",
          "value": "434"
        },
        {
          "name": "iterations",
          "value": "26"
        }
      ]
    },
    {
      "physicalTile": 175,
      "offset": 408342,
      "duration": 9602064,
      "phase": 11,
      "epoch": 0,
      "kind": "compute",
      "operation": "operation.24",
      "kernel": "repeat-remainder",
      "metadata": [
        {
          "name": "reason",
          "value": "Repeat"
        },
        {
          "name": "value",
          "value": "434"
        },
        {
          "name": "iterations",
          "value": "26"
        }
      ]
    }
  ]
}
