{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 189838,
  "matchedSampleCount": 97775,
  "profileSpanCycles": 10398354,
  "profileSpanMs": 6.932236,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 9836538,
      "phaseMs": 6.557692,
      "workCycles": 14474972136,
      "averageActiveTiles": 1471.5514885420052,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 9833540.85326087,
      "p50Cycles": 9833544,
      "p95Cycles": 9833556,
      "maximumCycles": 9833574,
      "firstOffset": 415932,
      "lastOffset": 10252470
    },
    {
      "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": 27120,
      "lastOffset": 10391844
    },
    {
      "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": 19716,
      "lastOffset": 10394652
    },
    {
      "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": 10390170
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 5993,
      "tileCount": 1440,
      "phaseCount": 122,
      "phaseCycles": 35268,
      "phaseMs": 0.023512,
      "workCycles": 18627216,
      "averageActiveTiles": 528.1619598502892,
      "usefulCycles": 678138.0,
      "estimatedWorkCycles": 18255372,
      "usefulUtilization": 0.03714731203505467,
      "workLaneOccupancy": 0.9058314053825891,
      "deviceUsefulUtilization": 0.013062594308919035,
      "meanCycles": 3108.1621892207577,
      "p50Cycles": 2088,
      "p95Cycles": 9300,
      "maximumCycles": 9300,
      "firstOffset": 0,
      "lastOffset": 10390152
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 30,
      "phaseCycles": 33084,
      "phaseMs": 0.022056,
      "workCycles": 23292360,
      "averageActiveTiles": 704.0369967355822,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23292360,
      "usefulUtilization": 0.6590315451074945,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.31520556370345837,
      "meanCycles": 29862.0,
      "p50Cycles": 29862,
      "p95Cycles": 29862,
      "maximumCycles": 29862,
      "firstOffset": 267690,
      "lastOffset": 300774
    },
    {
      "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": 35550,
      "lastOffset": 10267722
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 22,
      "phaseCycles": 30402,
      "phaseMs": 0.020268,
      "workCycles": 18315648,
      "averageActiveTiles": 602.4487862640616,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18315648,
      "usefulUtilization": 0.6520850367947669,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.26688032537905115,
      "meanCycles": 29352.0,
      "p50Cycles": 29352,
      "p95Cycles": 29352,
      "maximumCycles": 29352,
      "firstOffset": 269898,
      "lastOffset": 300300
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 17,
      "phaseCycles": 30114,
      "phaseMs": 0.020076,
      "workCycles": 807480,
      "averageActiveTiles": 26.814106395696353,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 807480,
      "usefulUtilization": 0.6580472581364244,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.01198705787572442,
      "meanCycles": 26916.0,
      "p50Cycles": 26916,
      "p95Cycles": 26916,
      "maximumCycles": 26916,
      "firstOffset": 267702,
      "lastOffset": 297816
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 34,
      "phaseCycles": 28662,
      "phaseMs": 0.019108,
      "workCycles": 32156190,
      "averageActiveTiles": 1121.9101946828553,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32156190,
      "usefulUtilization": 0.7348183973287881,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5600545184807638,
      "meanCycles": 26466.0,
      "p50Cycles": 26466,
      "p95Cycles": 26466,
      "maximumCycles": 26466,
      "firstOffset": 361200,
      "lastOffset": 389862
    },
    {
      "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": 148608,
      "lastOffset": 177198
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 14,
      "phaseCycles": 27498,
      "phaseMs": 0.018332,
      "workCycles": 635040,
      "averageActiveTiles": 23.094043203142046,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 635040,
      "usefulUtilization": 0.6510204081632653,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.01021378629908262,
      "meanCycles": 26460.0,
      "p50Cycles": 26460,
      "p95Cycles": 26460,
      "maximumCycles": 26460,
      "firstOffset": 269898,
      "lastOffset": 297396
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 21,
      "phaseCycles": 26676,
      "phaseMs": 0.017784,
      "workCycles": 6406452,
      "averageActiveTiles": 240.1578947368421,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6406452,
      "usefulUtilization": 0.7195221317509286,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.11739057090887754,
      "meanCycles": 26364.0,
      "p50Cycles": 26364,
      "p95Cycles": 26364,
      "maximumCycles": 26364,
      "firstOffset": 362766,
      "lastOffset": 389442
    },
    {
      "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": 49248,
      "lastOffset": 10342068
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 3,
      "phaseCycles": 19098,
      "phaseMs": 0.012732,
      "workCycles": 18143352,
      "averageActiveTiles": 950.0131950989633,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18143352,
      "usefulUtilization": 0.8366013071895425,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5399336147195017,
      "meanCycles": 18666.0,
      "p50Cycles": 18666,
      "p95Cycles": 18666,
      "maximumCycles": 18666,
      "firstOffset": 63564,
      "lastOffset": 82662
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 18810,
      "phaseMs": 0.01254,
      "workCycles": 9007524,
      "averageActiveTiles": 478.8688995215311,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 9007524,
      "usefulUtilization": 0.8322002805654473,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.27073018514666114,
      "meanCycles": 18534.0,
      "p50Cycles": 18534,
      "p95Cycles": 18534,
      "maximumCycles": 18534,
      "firstOffset": 63486,
      "lastOffset": 82296
    },
    {
      "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": 24402,
      "lastOffset": 10398354
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 7,
      "phaseCycles": 16800,
      "phaseMs": 0.0112,
      "workCycles": 13107420,
      "averageActiveTiles": 780.2035714285714,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 13107420,
      "usefulUtilization": 0.6486219255963416,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3437888198757764,
      "meanCycles": 16182.0,
      "p50Cycles": 16182,
      "p95Cycles": 16182,
      "maximumCycles": 16182,
      "firstOffset": 10277796,
      "lastOffset": 10294596
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 7,
      "phaseCycles": 16230,
      "phaseMs": 0.01082,
      "workCycles": 10318752,
      "averageActiveTiles": 635.7826247689464,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10318752,
      "usefulUtilization": 0.6410449635769907,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.2768785662621554,
      "meanCycles": 15924.0,
      "p50Cycles": 15924,
      "p95Cycles": 15924,
      "maximumCycles": 15924,
      "firstOffset": 10277970,
      "lastOffset": 10294200
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r81_r81",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 11,
      "phaseCycles": 12828,
      "phaseMs": 0.008552,
      "workCycles": 10839744,
      "averageActiveTiles": 845.0065481758653,
      "usefulCycles": 6298560.0,
      "estimatedWorkCycles": 10839744,
      "usefulUtilization": 0.5810616929698709,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.3335604181071298,
      "meanCycles": 12546.0,
      "p50Cycles": 12546,
      "p95Cycles": 12546,
      "maximumCycles": 12546,
      "firstOffset": 129714,
      "lastOffset": 142542
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r81_r81",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 12,
      "phaseCycles": 10698,
      "phaseMs": 0.007132,
      "workCycles": 5826816,
      "averageActiveTiles": 544.6640493550196,
      "usefulCycles": 3267378.0,
      "estimatedWorkCycles": 5826816,
      "usefulUtilization": 0.5607484430604982,
      "workLaneOccupancy": 0.875390625,
      "deviceUsefulUtilization": 0.20748608537150381,
      "meanCycles": 10116.0,
      "p50Cycles": 10116,
      "p95Cycles": 10116,
      "maximumCycles": 10116,
      "firstOffset": 129126,
      "lastOffset": 139824
    },
    {
      "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": 114882,
      "lastOffset": 10333242
    },
    {
      "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": 325668,
      "lastOffset": 10387338
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 20,
      "phaseCycles": 9768,
      "phaseMs": 0.006512,
      "workCycles": 9434880,
      "averageActiveTiles": 965.8968058968059,
      "usefulCycles": 4408992.0,
      "estimatedWorkCycles": 9434880,
      "usefulUtilization": 0.4673076923076923,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.30663791261617346,
      "meanCycles": 8190.0,
      "p50Cycles": 8190,
      "p95Cycles": 8190,
      "maximumCycles": 8190,
      "firstOffset": 184392,
      "lastOffset": 194160
    },
    {
      "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": 114000,
      "lastOffset": 10331844
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r56_r57",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 96,
      "phaseCount": 6,
      "phaseCycles": 8772,
      "phaseMs": 0.005848,
      "workCycles": 790848,
      "averageActiveTiles": 90.15595075239398,
      "usefulCycles": 373977.0,
      "estimatedWorkCycles": 790848,
      "usefulUtilization": 0.47288100873998545,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.028962660961755784,
      "meanCycles": 8238.0,
      "p50Cycles": 8238,
      "p95Cycles": 8238,
      "maximumCycles": 8238,
      "firstOffset": 185442,
      "lastOffset": 194214
    },
    {
      "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": 24396,
      "lastOffset": 10396986
    },
    {
      "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": 10335150,
      "lastOffset": 10343286
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 29,
      "phaseCycles": 7452,
      "phaseMs": 0.004968,
      "workCycles": 10324800,
      "averageActiveTiles": 1385.5072463768115,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10324800,
      "usefulUtilization": 0.732050209205021,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.6890359168241966,
      "meanCycles": 7170.0,
      "p50Cycles": 7170,
      "p95Cycles": 7170,
      "maximumCycles": 7170,
      "firstOffset": 219234,
      "lastOffset": 226686
    },
    {
      "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": 10351884,
      "lastOffset": 10359168
    },
    {
      "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": 199146,
      "lastOffset": 206124
    },
    {
      "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": 10344444,
      "lastOffset": 10350564
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 3,
      "phaseCycles": 4380,
      "phaseMs": 0.00292,
      "workCycles": 1449360,
      "averageActiveTiles": 330.90410958904107,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1449360,
      "usefulUtilization": 0.671833084947839,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.15102739726027398,
      "meanCycles": 4026.0,
      "p50Cycles": 4026,
      "p95Cycles": 4026,
      "maximumCycles": 4026,
      "firstOffset": 12372,
      "lastOffset": 16752
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 2,
      "phaseCycles": 4326,
      "phaseMs": 0.002884,
      "workCycles": 4322160,
      "averageActiveTiles": 999.1123439667128,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4322160,
      "usefulUtilization": 0.6672913543228386,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.4529205360911777,
      "meanCycles": 4002.0,
      "p50Cycles": 4002,
      "p95Cycles": 4002,
      "maximumCycles": 4002,
      "firstOffset": 12378,
      "lastOffset": 16704
    },
    {
      "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": 10381764,
      "lastOffset": 10385070
    },
    {
      "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": 10381278,
      "lastOffset": 10383816
    },
    {
      "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": 10373538,
      "lastOffset": 10375878
    },
    {
      "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": 10376154,
      "lastOffset": 10378350
    },
    {
      "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": 10391232,
      "lastOffset": 10393392
    },
    {
      "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": 10316262,
      "lastOffset": 10369236
    },
    {
      "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": 10361946,
      "lastOffset": 10363560
    }
  ],
  "samples": []
}
