{
  "schema_version": 1,
  "precision_comparison": {
    "class": "same_storage_precision",
    "same_weight_representation": true,
    "baseline_weights": "float16",
    "optimized_weights": "float16",
    "bitwise_exact": false,
    "validation": "float32_accumulation_float16_output_tolerance_checked"
  },
  "recorded_at": "2026-10-02T05:57:54.421312+00:00",
  "scope": "isolated_primitive",
  "hardware": {
    "chip": "Apple M5",
    "system": "Darwin",
    "macos": "26.2",
    "machine": "arm64"
  },
  "software": {
    "python": "3.12.12",
    "mlx": "0.32.0",
    "metal_compiler": "Apple metal version 32023.864 (metalfe-32023.864)"
  },
  "implementation": {
    "baseline": "_lower_gemm (SIMDgroup matrices)",
    "candidate": "_lower_tensor_ops_gemm (generic MPP cooperative tensors)",
    "input_dtype": "float16",
    "accumulator_dtype": "float32",
    "output_dtype": "float16",
    "relaxed_precision": false,
    "global_fast_math": true,
    "block_m": 64,
    "block_n": 64,
    "block_k": 16,
    "wm": 2,
    "wn": 2,
    "swizzle": "auto"
  },
  "validation": {
    "reference": "float32 NumPy matmul of float16 inputs, rounded to float16",
    "rtol": 0.002,
    "atol": 0.005,
    "bitwise_exact": false
  },
  "methodology": {
    "seed": "943 + square dimension",
    "resident_buffers": true,
    "compilation_excluded": true,
    "source_checks": [
      "half, half, float cooperative tensor types",
      "false relaxed_precision in matmul2d descriptor"
    ],
    "paired_wall_rounds": 4,
    "paired_wall_round_ms": 180,
    "paired_order": "Interleaved alternating AB/BA within every round; caller ordering reversed on odd rounds",
    "gpu_samples_per_variant": 50,
    "gpu_warmup_pairs": 10,
    "gpu_measurement": "Metal command-buffer GPU elapsed time after synchronized alternating dispatches",
    "mlx_wall_rounds": 3,
    "mlx_work": "mx.eval(left @ right), inputs resident",
    "notes": [
      "The first 256-square wall round has a visibly different clock/latency regime; retain it rather than selectively deleting it.",
      "GPU elapsed time is compared only between the two meTile variants, not against MLX wall-clock latency.",
      "MLX 1024-square wall latency remains lower than the candidate."
    ]
  },
  "cases": [
    {
      "shape": [
        256,
        256,
        256
      ],
      "max_absolute_error": 0.015625,
      "wall_rounds_microseconds": [
        {
          "baseline": 269.958,
          "candidate": 106.125
        },
        {
          "baseline": 486.5,
          "candidate": 326.25000000000006
        },
        {
          "baseline": 496.3330000000001,
          "candidate": 335.708
        },
        {
          "baseline": 488.9580000000001,
          "candidate": 326.854
        }
      ],
      "gpu_median_microseconds": {
        "baseline": 199.37499018851668,
        "candidate": 30.14583489857614
      },
      "mlx_wall_rounds_microseconds": [
        {
          "candidate": 222.25,
          "mlx": 332.417
        },
        {
          "candidate": 329.542,
          "mlx": 358.167
        },
        {
          "candidate": 322.8955,
          "mlx": 351.29150000000004
        }
      ]
    },
    {
      "shape": [
        1024,
        1024,
        1024
      ],
      "max_absolute_error": 0.0625,
      "wall_rounds_microseconds": [
        {
          "baseline": 1042.3960000000002,
          "candidate": 374.437
        },
        {
          "baseline": 1024.2085000000002,
          "candidate": 377.0830000000001
        },
        {
          "baseline": 1028.0420000000001,
          "candidate": 375.66600000000005
        },
        {
          "baseline": 1035.1045,
          "candidate": 377.41700000000003
        }
      ],
      "gpu_median_microseconds": {
        "baseline": 833.3541773026809,
        "candidate": 161.2083287909627
      },
      "mlx_wall_rounds_microseconds": [
        {
          "candidate": 376.83349999999996,
          "mlx": 359.66650000000004
        },
        {
          "candidate": 378.97850000000005,
          "mlx": 361.60400000000004
        },
        {
          "candidate": 379.083,
          "mlx": 359.04200000000003
        }
      ]
    }
  ]
}