{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 30741,
  "matchedSampleCount": 29346,
  "profileSpanCycles": 333738,
  "profileSpanMs": 0.222492,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 86,
      "tileCount": 16,
      "phaseCount": 6,
      "phaseCycles": 326160,
      "phaseMs": 0.21744,
      "workCycles": 4834524,
      "averageActiveTiles": 14.822553348050036,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 56215.395348837206,
      "p50Cycles": 35370,
      "p95Cycles": 138528,
      "maximumCycles": 139152,
      "firstOffset": 6372,
      "lastOffset": 332532
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 10218,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 119154,
      "phaseMs": 0.079436,
      "workCycles": 130362210,
      "averageActiveTiles": 1094.06490759857,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 12758.094539048738,
      "p50Cycles": 7560,
      "p95Cycles": 32424,
      "maximumCycles": 32688,
      "firstOffset": 0,
      "lastOffset": 333108
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k288_c336_r52",
      "dimensions": {},
      "sampleCount": 1352,
      "tileCount": 1352,
      "phaseCount": 4,
      "phaseCycles": 74142,
      "phaseMs": 0.049428,
      "workCycles": 93377232,
      "averageActiveTiles": 1259.4377276037874,
      "usefulCycles": 52371072.0,
      "estimatedWorkCycles": 93377232,
      "usefulUtilization": 0.5608548345061246,
      "workLaneOccupancy": 0.9853479853479854,
      "deviceUsefulUtilization": 0.47986531133559224,
      "meanCycles": 69066.0,
      "p50Cycles": 69066,
      "p95Cycles": 69066,
      "maximumCycles": 69066,
      "firstOffset": 71136,
      "lastOffset": 145278
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k288_c336_r53",
      "dimensions": {},
      "sampleCount": 104,
      "tileCount": 104,
      "phaseCount": 4,
      "phaseCycles": 74106,
      "phaseMs": 0.049404,
      "workCycles": 7183488,
      "averageActiveTiles": 96.93530888187192,
      "usefulCycles": 4106016.0,
      "estimatedWorkCycles": 7183488,
      "usefulUtilization": 0.5715908483455391,
      "workLaneOccupancy": 0.9853479853479854,
      "deviceUsefulUtilization": 0.03764085287936551,
      "meanCycles": 69072.0,
      "p50Cycles": 69072,
      "p95Cycles": 69072,
      "maximumCycles": 69072,
      "firstOffset": 71190,
      "lastOffset": 145296
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k160_c384_r81",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 6,
      "phaseCycles": 58446,
      "phaseMs": 0.038964,
      "workCycles": 84575664,
      "averageActiveTiles": 1447.073606405913,
      "usefulCycles": 56477088.0,
      "estimatedWorkCycles": 84575664,
      "usefulUtilization": 0.6677699627637568,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.6564621523546819,
      "meanCycles": 58008.0,
      "p50Cycles": 58008,
      "p95Cycles": 58008,
      "maximumCycles": 58008,
      "firstOffset": 238476,
      "lastOffset": 296922
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r52_p52_c288_p288",
      "dimensions": {},
      "sampleCount": 1352,
      "tileCount": 1352,
      "phaseCount": 4,
      "phaseCycles": 48612,
      "phaseMs": 0.032408,
      "workCycles": 58860672,
      "averageActiveTiles": 1210.8259688965688,
      "usefulCycles": 5061888.0,
      "estimatedWorkCycles": 58860672,
      "usefulUtilization": 0.08599779492833518,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07073937728741159,
      "meanCycles": 43536.0,
      "p50Cycles": 43536,
      "p95Cycles": 43536,
      "maximumCycles": 43536,
      "firstOffset": 27600,
      "lastOffset": 76212
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r53_p53_c288_p288",
      "dimensions": {},
      "sampleCount": 104,
      "tileCount": 104,
      "phaseCount": 4,
      "phaseCycles": 48570,
      "phaseMs": 0.03238,
      "workCycles": 4527744,
      "averageActiveTiles": 93.22100061766523,
      "usefulCycles": 396864.0,
      "estimatedWorkCycles": 4527744,
      "usefulUtilization": 0.087651598676957,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.005550930526089642,
      "meanCycles": 43536.0,
      "p50Cycles": 43536,
      "p95Cycles": 43536,
      "maximumCycles": 43536,
      "firstOffset": 27654,
      "lastOffset": 76224
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2914,
      "tileCount": 1458,
      "phaseCount": 10,
      "phaseCycles": 35016,
      "phaseMs": 0.023344,
      "workCycles": 41462484,
      "averageActiveTiles": 1184.1010966415354,
      "usefulCycles": 15623928.0,
      "estimatedWorkCycles": 41462484,
      "usefulUtilization": 0.376820838809368,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3031209024942635,
      "meanCycles": 14228.717913520934,
      "p50Cycles": 13818,
      "p95Cycles": 14856,
      "maximumCycles": 14856,
      "firstOffset": 154440,
      "lastOffset": 332286
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 6,
      "phaseCycles": 22458,
      "phaseMs": 0.014972,
      "workCycles": 30792960,
      "averageActiveTiles": 1371.1354528453112,
      "usefulCycles": 18825696.0,
      "estimatedWorkCycles": 30792960,
      "usefulUtilization": 0.6113636363636363,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5694717095099258,
      "meanCycles": 21120.0,
      "p50Cycles": 21120,
      "p95Cycles": 21120,
      "maximumCycles": 21120,
      "firstOffset": 173448,
      "lastOffset": 195906
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c144_p160",
      "dimensions": {},
      "sampleCount": 54,
      "tileCount": 54,
      "phaseCount": 5,
      "phaseCycles": 11868,
      "phaseMs": 0.007912,
      "workCycles": 629532,
      "averageActiveTiles": 53.04448938321537,
      "usefulCycles": 314928.0,
      "estimatedWorkCycles": 629532,
      "usefulUtilization": 0.5002573340195574,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.018027102475051654,
      "meanCycles": 11658.0,
      "p50Cycles": 11658,
      "p95Cycles": 11658,
      "maximumCycles": 11658,
      "firstOffset": 203106,
      "lastOffset": 214974
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c160_p160",
      "dimensions": {},
      "sampleCount": 1404,
      "tileCount": 1404,
      "phaseCount": 5,
      "phaseCycles": 11178,
      "phaseMs": 0.007452,
      "workCycles": 15188472,
      "averageActiveTiles": 1358.7826086956522,
      "usefulCycles": 9097920.0,
      "estimatedWorkCycles": 15188472,
      "usefulUtilization": 0.5990016638935108,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.552930056710775,
      "meanCycles": 10818.0,
      "p50Cycles": 10818,
      "p95Cycles": 10818,
      "maximumCycles": 10818,
      "firstOffset": 202992,
      "lastOffset": 214170
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 5662,
      "tileCount": 1458,
      "phaseCount": 17,
      "phaseCycles": 10146,
      "phaseMs": 0.006764,
      "workCycles": 7338516,
      "averageActiveTiles": 723.2915434654051,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1296.0996114447191,
      "p50Cycles": 1014,
      "p95Cycles": 2934,
      "maximumCycles": 3126,
      "firstOffset": 152160,
      "lastOffset": 333738
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 6,
      "phaseCycles": 8976,
      "phaseMs": 0.005984,
      "workCycles": 11389896,
      "averageActiveTiles": 1268.927807486631,
      "usefulCycles": 2353212.0,
      "estimatedWorkCycles": 11389896,
      "usefulUtilization": 0.206605222734255,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.17810265781213672,
      "meanCycles": 7812.0,
      "p50Cycles": 7812,
      "p95Cycles": 7812,
      "maximumCycles": 7812,
      "firstOffset": 213810,
      "lastOffset": 222786
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 1587,
      "tileCount": 1458,
      "phaseCount": 9,
      "phaseCycles": 4644,
      "phaseMs": 0.003096,
      "workCycles": 511092,
      "averageActiveTiles": 110.05426356589147,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 322.04914933837426,
      "p50Cycles": 330,
      "p95Cycles": 330,
      "maximumCycles": 330,
      "firstOffset": 0,
      "lastOffset": 196236
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 92,
      "tileCount": 85,
      "phaseCount": 3,
      "phaseCycles": 2676,
      "phaseMs": 0.001784,
      "workCycles": 42978,
      "averageActiveTiles": 16.06053811659193,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 467.1521739130435,
      "p50Cycles": 144,
      "p95Cycles": 1554,
      "maximumCycles": 1758,
      "firstOffset": 0,
      "lastOffset": 2676
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 43,
      "tileCount": 43,
      "phaseCount": 1,
      "phaseCycles": 2484,
      "phaseMs": 0.001656,
      "workCycles": 53220,
      "averageActiveTiles": 21.42512077294686,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1237.6744186046512,
      "p50Cycles": 1242,
      "p95Cycles": 2238,
      "maximumCycles": 2484,
      "firstOffset": 0,
      "lastOffset": 2484
    }
  ],
  "samples": []
}
