outputs/nvfp4/kernel_speed.json

{
  "T": 1536,
  "per_step_ms": {
    "bf16": 73.842,
    "nvfp4_w4a4_r64": 26.82,
    "nvfp4_w4a4_r128": 29.674,
    "fp8_r64": 60.331,
    "fp8_r128": 60.8
  },
  "speedup_vs_bf16": {
    "bf16": 1.0,
    "nvfp4_w4a4_r64": 2.753,
    "nvfp4_w4a4_r128": 2.488,
    "fp8_r64": 1.224,
    "fp8_r128": 1.215
  },
  "per_shape_us": {
    "3072x27648x20": {
      "bf16": 1752.2,
      "nvfp4_w4a4_r64": 571.6,
      "nvfp4_w4a4_r128": 626.1,
      "fp8_r64": 1431.2,
      "fp8_r128": 1400.8
    },
    "12288x3072x20": {
      "bf16": 717.6,
      "nvfp4_w4a4_r64": 303.9,
      "nvfp4_w4a4_r128": 339.1,
      "fp8_r64": 625.1,
      "fp8_r128": 651.4
    },
    "3072x18432x10": {
      "bf16": 1156.0,
      "nvfp4_w4a4_r64": 403.1,
      "nvfp4_w4a4_r128": 432.6,
      "fp8_r64": 953.9,
      "fp8_r128": 934.5
    },
    "3072x3072x40": {
      "bf16": 186.8,
      "nvfp4_w4a4_r64": 79.2,
      "nvfp4_w4a4_r128": 89.3,
      "fp8_r64": 139.6,
      "fp8_r128": 154.1
    },
    "9216x3072x10": {
      "bf16": 541.2,
      "nvfp4_w4a4_r64": 211.1,
      "nvfp4_w4a4_r128": 247.2,
      "fp8_r64": 408.1,
      "fp8_r128": 424.6
    }
  }
}