{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 41024,
  "matchedSampleCount": 40854,
  "profileSpanCycles": 275562,
  "profileSpanMs": 0.183708,
  "atOffset": null,
  "groups": [
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 8800,
      "tileCount": 1472,
      "phaseCount": 6,
      "phaseCycles": 121248,
      "phaseMs": 0.080832,
      "workCycles": 122044308,
      "averageActiveTiles": 1006.5675969912905,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 13868.671363636364,
      "p50Cycles": 10134,
      "p95Cycles": 35802,
      "maximumCycles": 37482,
      "firstOffset": 0,
      "lastOffset": 275094
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 79308,
      "phaseMs": 0.052872,
      "workCycles": 2514426,
      "averageActiveTiles": 31.70456952640339,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 78575.8125,
      "p50Cycles": 78516,
      "p95Cycles": 79212,
      "maximumCycles": 79308,
      "firstOffset": 15258,
      "lastOffset": 94566
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k192_c288_r46",
      "dimensions": {},
      "sampleCount": 828,
      "tileCount": 828,
      "phaseCount": 9,
      "phaseCycles": 77694,
      "phaseMs": 0.051796,
      "workCycles": 60862968,
      "averageActiveTiles": 783.3676731793961,
      "usefulCycles": 32073408.0,
      "estimatedWorkCycles": 60862968,
      "usefulUtilization": 0.5269773896008489,
      "workLaneOccupancy": 0.9746376811594203,
      "deviceUsefulUtilization": 0.28044636651478877,
      "meanCycles": 73506.0,
      "p50Cycles": 73506,
      "p95Cycles": 73506,
      "maximumCycles": 73506,
      "firstOffset": 178098,
      "lastOffset": 255792
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k192_c288_r45",
      "dimensions": {},
      "sampleCount": 644,
      "tileCount": 644,
      "phaseCount": 13,
      "phaseCycles": 76140,
      "phaseMs": 0.05076,
      "workCycles": 46507104,
      "averageActiveTiles": 610.8104018912529,
      "usefulCycles": 24403680.0,
      "estimatedWorkCycles": 46507104,
      "usefulUtilization": 0.5247301573540248,
      "workLaneOccupancy": 0.9746376811594203,
      "deviceUsefulUtilization": 0.2177382053654024,
      "meanCycles": 72216.0,
      "p50Cycles": 72216,
      "p95Cycles": 72216,
      "maximumCycles": 72216,
      "firstOffset": 178152,
      "lastOffset": 254292
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k192_c64_r216",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 3,
      "phaseCycles": 64314,
      "phaseMs": 0.042876,
      "workCycles": 42788736,
      "averageActiveTiles": 665.3098236775819,
      "usefulCycles": 34012224.0,
      "estimatedWorkCycles": 42788736,
      "usefulUtilization": 0.7948873273564332,
      "workLaneOccupancy": 0.94921875,
      "deviceUsefulUtilization": 0.35927061658087833,
      "meanCycles": 49524.0,
      "p50Cycles": 49524,
      "p95Cycles": 49524,
      "maximumCycles": 49524,
      "firstOffset": 30000,
      "lastOffset": 94314
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k192_c64_r214",
      "dimensions": {},
      "sampleCount": 576,
      "tileCount": 576,
      "phaseCount": 4,
      "phaseCycles": 64152,
      "phaseMs": 0.042768,
      "workCycles": 28162944,
      "averageActiveTiles": 439.003367003367,
      "usefulCycles": 22464864.0,
      "estimatedWorkCycles": 28162944,
      "usefulUtilization": 0.7976745612958646,
      "workLaneOccupancy": 0.94921875,
      "deviceUsefulUtilization": 0.23789525691699603,
      "meanCycles": 48894.0,
      "p50Cycles": 48894,
      "p95Cycles": 48894,
      "maximumCycles": 48894,
      "firstOffset": 29466,
      "lastOffset": 93618
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 15316,
      "tileCount": 1472,
      "phaseCount": 34,
      "phaseCycles": 28596,
      "phaseMs": 0.019064,
      "workCycles": 32249448,
      "averageActiveTiles": 1127.760805707092,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2105.6051188299816,
      "p50Cycles": 546,
      "p95Cycles": 5076,
      "maximumCycles": 5076,
      "firstOffset": 18960,
      "lastOffset": 275562
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2912,
      "tileCount": 1472,
      "phaseCount": 20,
      "phaseCycles": 20724,
      "phaseMs": 0.013816,
      "workCycles": 22761408,
      "averageActiveTiles": 1098.3115228720324,
      "usefulCycles": 8540964.0,
      "estimatedWorkCycles": 22761408,
      "usefulUtilization": 0.3752388252958692,
      "workLaneOccupancy": 0.9956100185764057,
      "deviceUsefulUtilization": 0.2799790255784094,
      "meanCycles": 7816.417582417583,
      "p50Cycles": 9510,
      "p95Cycles": 9510,
      "maximumCycles": 9510,
      "firstOffset": 100254,
      "lastOffset": 273996
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 6339,
      "tileCount": 1457,
      "phaseCount": 25,
      "phaseCycles": 15138,
      "phaseMs": 0.010092,
      "workCycles": 3876144,
      "averageActiveTiles": 256.05390408244153,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 611.4756270705159,
      "p50Cycles": 270,
      "p95Cycles": 1866,
      "maximumCycles": 1866,
      "firstOffset": 0,
      "lastOffset": 275016
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 8,
      "phaseCycles": 13596,
      "phaseMs": 0.009064,
      "workCycles": 16898664,
      "averageActiveTiles": 1242.9143865842896,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 16898664,
      "usefulUtilization": 0.5570172884673014,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4703293487854484,
      "meanCycles": 11480.070652173914,
      "p50Cycles": 11370,
      "p95Cycles": 11874,
      "maximumCycles": 12354,
      "firstOffset": 123246,
      "lastOffset": 136842
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r64_p64_c36_p36",
      "dimensions": {},
      "sampleCount": 1120,
      "tileCount": 1120,
      "phaseCount": 2,
      "phaseCycles": 11934,
      "phaseMs": 0.007956,
      "workCycles": 11854080,
      "averageActiveTiles": 993.3031674208145,
      "usefulCycles": 1290240.0,
      "estimatedWorkCycles": 11854080,
      "usefulUtilization": 0.10884353741496598,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07344743917633943,
      "meanCycles": 10584.0,
      "p50Cycles": 10584,
      "p95Cycles": 10584,
      "maximumCycles": 10584,
      "firstOffset": 108726,
      "lastOffset": 120660
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r46_p46_c80_p192",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 7,
      "phaseCycles": 11790,
      "phaseMs": 0.00786,
      "workCycles": 374760,
      "averageActiveTiles": 31.786259541984734,
      "usefulCycles": 66240.0,
      "estimatedWorkCycles": 374760,
      "usefulUtilization": 0.17675312199807877,
      "workLaneOccupancy": 0.4166666666666667,
      "deviceUsefulUtilization": 0.003816793893129771,
      "meanCycles": 10410.0,
      "p50Cycles": 10410,
      "p95Cycles": 10410,
      "maximumCycles": 10410,
      "firstOffset": 170496,
      "lastOffset": 182286
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r41_p48_c36_p36",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 1,
      "phaseCycles": 11424,
      "phaseMs": 0.007616,
      "workCycles": 1101408,
      "averageActiveTiles": 96.41176470588235,
      "usefulCycles": 82656.0,
      "estimatedWorkCycles": 1101408,
      "usefulUtilization": 0.075045759609518,
      "workLaneOccupancy": 0.8541666666666666,
      "deviceUsefulUtilization": 0.004915281329923273,
      "meanCycles": 9834.0,
      "p50Cycles": 9834,
      "p95Cycles": 9834,
      "maximumCycles": 9834,
      "firstOffset": 105096,
      "lastOffset": 116520
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r45_p45_c80_p192",
      "dimensions": {},
      "sampleCount": 28,
      "tileCount": 28,
      "phaseCount": 8,
      "phaseCycles": 10860,
      "phaseMs": 0.00724,
      "workCycles": 291480,
      "averageActiveTiles": 26.83977900552486,
      "usefulCycles": 50400.0,
      "estimatedWorkCycles": 291480,
      "usefulUtilization": 0.1729106628242075,
      "workLaneOccupancy": 0.4166666666666667,
      "deviceUsefulUtilization": 0.0031527744415085273,
      "meanCycles": 10410.0,
      "p50Cycles": 10410,
      "p95Cycles": 10410,
      "maximumCycles": 10410,
      "firstOffset": 171216,
      "lastOffset": 182076
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r64_p64_c34_p34",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 1,
      "phaseCycles": 10704,
      "phaseMs": 0.007136,
      "workCycles": 803520,
      "averageActiveTiles": 75.06726457399103,
      "usefulCycles": 87040.0,
      "estimatedWorkCycles": 803520,
      "usefulUtilization": 0.10832337714058145,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.005524143757717554,
      "meanCycles": 10044.0,
      "p50Cycles": 10044,
      "p95Cycles": 10044,
      "maximumCycles": 10044,
      "firstOffset": 108444,
      "lastOffset": 119148
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r41_p48_c34_p34",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 2,
      "phaseCycles": 9642,
      "phaseMs": 0.006428,
      "workCycles": 74592,
      "averageActiveTiles": 7.736154324828874,
      "usefulCycles": 5576.0,
      "estimatedWorkCycles": 74592,
      "usefulUtilization": 0.07475332475332476,
      "workLaneOccupancy": 0.8541666666666666,
      "deviceUsefulUtilization": 0.00039286906018055066,
      "meanCycles": 9324.0,
      "p50Cycles": 9324,
      "p95Cycles": 9324,
      "maximumCycles": 9324,
      "firstOffset": 105936,
      "lastOffset": 115578
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r46_p46_c192_p192",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 9,
      "phaseCycles": 8574,
      "phaseMs": 0.005716,
      "workCycles": 5811696,
      "averageActiveTiles": 677.8278516445066,
      "usefulCycles": 3497472.0,
      "estimatedWorkCycles": 5811696,
      "usefulUtilization": 0.6017988552739166,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.27711686494051785,
      "meanCycles": 7338.0,
      "p50Cycles": 7338,
      "p95Cycles": 7338,
      "maximumCycles": 7338,
      "firstOffset": 170760,
      "lastOffset": 179334
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r45_p45_c192_p192",
      "dimensions": {},
      "sampleCount": 616,
      "tileCount": 616,
      "phaseCount": 13,
      "phaseCycles": 8514,
      "phaseMs": 0.005676,
      "workCycles": 4520208,
      "averageActiveTiles": 530.9147286821706,
      "usefulCycles": 2661120.0,
      "estimatedWorkCycles": 4520208,
      "usefulUtilization": 0.588716271463614,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2123356926188069,
      "meanCycles": 7338.0,
      "p50Cycles": 7338,
      "p95Cycles": 7338,
      "maximumCycles": 7338,
      "firstOffset": 170814,
      "lastOffset": 179328
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r48_p48_c36_p36",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 2,
      "phaseCycles": 7842,
      "phaseMs": 0.005228,
      "workCycles": 799680,
      "averageActiveTiles": 101.97398622800306,
      "usefulCycles": 96768.0,
      "estimatedWorkCycles": 799680,
      "usefulUtilization": 0.12100840336134454,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.008382954658860317,
      "meanCycles": 7140.0,
      "p50Cycles": 7140,
      "p95Cycles": 7140,
      "maximumCycles": 7140,
      "firstOffset": 106878,
      "lastOffset": 114720
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r48_p48_c34_p34",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 2,
      "phaseCycles": 7080,
      "phaseMs": 0.00472,
      "workCycles": 54240,
      "averageActiveTiles": 7.661016949152542,
      "usefulCycles": 6528.0,
      "estimatedWorkCycles": 54240,
      "usefulUtilization": 0.12035398230088495,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0006263817243920412,
      "meanCycles": 6780.0,
      "p50Cycles": 6780,
      "p95Cycles": 6780,
      "maximumCycles": 6780,
      "firstOffset": 106602,
      "lastOffset": 113682
    }
  ],
  "samples": []
}
