{
  "status": "four_pairs_audited",
  "generated_at": "2026-09-10T05:42:14.477258+00:00",
  "rows": [
    {
      "case": "480_short",
      "audit": "analysis/attention_hardware_480_short_audit_v1.json",
      "audit_sha256": "6267d155da94bb023f6a337e3b72c833d3d5ea3f077bbaef86c8c821dd810071",
      "valid_keys": 4680,
      "versions": [
        {
          "version": 2,
          "total_ncu_kernel_us": 332.48,
          "kernel_count": 1,
          "fp64_sample_relative_l2": 0.0019169673843473084,
          "full_output_vs_capture": {
            "exact": false,
            "allclose": true,
            "relative_l2": 0.0009067576358098216,
            "max_abs": 0.03125
          }
        },
        {
          "version": 3,
          "total_ncu_kernel_us": 179.744,
          "kernel_count": 2,
          "fp64_sample_relative_l2": 0.0019408478232929383,
          "full_output_vs_capture": {
            "exact": true,
            "allclose": true,
            "relative_l2": 0.0,
            "max_abs": 0.0
          }
        }
      ],
      "fa2_over_fa3_ncu_duration_ratio": 1.8497418550827844
    },
    {
      "case": "704_short",
      "audit": "analysis/attention_hardware_704_short_audit_v1.json",
      "audit_sha256": "c27a41e71adf40bdccfc55a0ced42afc4f98d341365bc6f7eec8df6917d3a37b",
      "valid_keys": 10560,
      "versions": [
        {
          "version": 2,
          "total_ncu_kernel_us": 1594.7839999999999,
          "kernel_count": 1,
          "fp64_sample_relative_l2": 0.001885860804486142,
          "full_output_vs_capture": {
            "exact": false,
            "allclose": true,
            "relative_l2": 0.0008376689748025244,
            "max_abs": 0.03125
          }
        },
        {
          "version": 3,
          "total_ncu_kernel_us": 821.184,
          "kernel_count": 2,
          "fp64_sample_relative_l2": 0.0019240266183567505,
          "full_output_vs_capture": {
            "exact": true,
            "allclose": true,
            "relative_l2": 0.0,
            "max_abs": 0.0
          }
        }
      ],
      "fa2_over_fa3_ncu_duration_ratio": 1.9420543995012078
    },
    {
      "case": "704_long",
      "audit": "analysis/attention_hardware_704_long_audit_v1.json",
      "audit_sha256": "37098c366485c99cf8620310e633c72011438f54cf6babf115a5d98969c64cb7",
      "valid_keys": 63360,
      "versions": [
        {
          "version": 2,
          "total_ncu_kernel_us": 9490.591999999999,
          "kernel_count": 1,
          "fp64_sample_relative_l2": 0.0017342987872260081,
          "full_output_vs_capture": {
            "exact": false,
            "allclose": true,
            "relative_l2": 0.0004884921109147934,
            "max_abs": 0.0625
          }
        },
        {
          "version": 3,
          "total_ncu_kernel_us": 4683.872,
          "kernel_count": 2,
          "fp64_sample_relative_l2": 0.0017431127374990974,
          "full_output_vs_capture": {
            "exact": true,
            "allclose": true,
            "relative_l2": 0.0,
            "max_abs": 0.0
          }
        }
      ],
      "fa2_over_fa3_ncu_duration_ratio": 2.026227873007631
    },
    {
      "case": "704_cross",
      "audit": "analysis/attention_hardware_704_cross_audit_v1.json",
      "audit_sha256": "a240d262d4e923f224e933b658b29686b466621076830d0028e0bca29f6ae0af",
      "valid_keys": 512,
      "versions": [
        {
          "version": 2,
          "total_ncu_kernel_us": 92.928,
          "kernel_count": 1,
          "fp64_sample_relative_l2": 0.0018472856840454418,
          "full_output_vs_capture": {
            "exact": false,
            "allclose": true,
            "relative_l2": 0.000302717243891442,
            "max_abs": 0.00048828125
          }
        },
        {
          "version": 3,
          "total_ncu_kernel_us": 80.41600000000001,
          "kernel_count": 2,
          "fp64_sample_relative_l2": 0.001852340011960572,
          "full_output_vs_capture": {
            "exact": false,
            "allclose": true,
            "relative_l2": 0.00011005898194572147,
            "max_abs": 0.00048828125
          }
        }
      ],
      "fa2_over_fa3_ncu_duration_ratio": 1.1555909271786706
    }
  ],
  "scope": "同一分配GPU0；NCU内核耗时之和，FA3包括准备kernel；非未插桩延迟、非整模型加速。FP64检查为32个query、4个head及全部有效KV；另保存并比较完整输出。"
}