﻿{
  "schema": "fak.strix.validation/v1",
  "timestamp": "2026-09-07T05:55:53Z",
  "verdict": "PASS",
  "target": {
    "mode": "ssh",
    "host": "strix1",
    "reachable": true,
    "cpu_model": "AMD RYZEN AI MAX+ 395 w/ Radeon 8060S",
    "gpu_name": "AMD Radeon 8060S Graphics (RADV STRIX_HALO)",
    "target_isa": "gfx1151",
    "compute_units": 40,
    "total_ram_bytes": 67028504576,
    "uma_buffer_bytes": 58985084026,
    "power_dpm_level": "manual",
    "lockup_timeout": -1,
    "vulkan_icd": "/usr/share/vulkan/icd.d/radeon_icd.json",
    "roundtrip_latency_ms": 442.6651,
    "discovered_at": "2026-09-07T05:54:57Z"
  },
  "provenance": {
    "git_ref": "HEAD",
    "command": "fak-dev amd-strix-validate --subkernels=all --ablate=all --json",
    "generated_by": "fak/internal/amdgpu",
    "transport": "ssh"
  },
  "subkernels": [
    {
      "name": "argmax",
      "status": "PASS",
      "duration_us": 829863,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (argmax)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": true,
        "passed": true
      },
      "metrics": {
        "category": "reduction",
        "wall_ms": 829
      }
    },
    {
      "name": "matmul_f32",
      "status": "PASS",
      "duration_us": 780377,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (matmul_f32)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "gemv",
        "wall_ms": 780
      }
    },
    {
      "name": "matmul2_f32",
      "status": "PASS",
      "duration_us": 618973,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (matmul2_f32)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "gemv",
        "wall_ms": 618
      }
    },
    {
      "name": "matmul3_f32",
      "status": "PASS",
      "duration_us": 555106,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (matmul3_f32)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "gemv",
        "wall_ms": 555
      }
    },
    {
      "name": "q8_matmul",
      "status": "PASS",
      "duration_us": 613932,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (q8_matmul)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "quant",
        "wall_ms": 613
      }
    },
    {
      "name": "q8_matmul_wide",
      "status": "PASS",
      "duration_us": 597134,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (q8_matmul_wide)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "quant",
        "wall_ms": 597
      }
    },
    {
      "name": "q8_matmul_vocab",
      "status": "PASS",
      "duration_us": 819693,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (q8_matmul_vocab)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "quant",
        "wall_ms": 819
      }
    },
    {
      "name": "q4k_matmul",
      "status": "PASS",
      "duration_us": 715857,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (q4k_matmul)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "quant",
        "wall_ms": 715
      }
    },
    {
      "name": "q2k_matmul",
      "status": "PASS",
      "duration_us": 1127285,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (q2k_matmul)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "quant",
        "wall_ms": 1127
      }
    },
    {
      "name": "rmsnorm",
      "status": "PASS",
      "duration_us": 564887,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (rmsnorm)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "norm",
        "wall_ms": 564
      }
    },
    {
      "name": "rmsnorm_matmul",
      "status": "PASS",
      "duration_us": 586984,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (rmsnorm_matmul)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "fused",
        "wall_ms": 586
      }
    },
    {
      "name": "rmsnorm_matmul2",
      "status": "PASS",
      "duration_us": 592157,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (rmsnorm_matmul2)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "fused",
        "wall_ms": 592
      }
    },
    {
      "name": "rmsnorm_matmul3",
      "status": "PASS",
      "duration_us": 614689,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (rmsnorm_matmul3)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "fused",
        "wall_ms": 614
      }
    },
    {
      "name": "swiglu",
      "status": "PASS",
      "duration_us": 562894,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (swiglu)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "activation",
        "wall_ms": 562
      }
    },
    {
      "name": "swiglu_matmul_add",
      "status": "PASS",
      "duration_us": 717119,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (swiglu_matmul_add)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "fused",
        "wall_ms": 717
      }
    },
    {
      "name": "rope",
      "status": "PASS",
      "duration_us": 624330,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (rope)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "positional",
        "wall_ms": 624
      }
    },
    {
      "name": "attention",
      "status": "PASS",
      "duration_us": 611909,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (attention)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "attention",
        "wall_ms": 611
      }
    },
    {
      "name": "qwen35_gdn_decode",
      "status": "PASS",
      "duration_us": 579486,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (qwen35_gdn_decode)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "linear_attention",
        "wall_ms": 579
      }
    },
    {
      "name": "qwen35_gdn_preprojected",
      "status": "PASS",
      "duration_us": 598055,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (qwen35_gdn_preprojected)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "linear_attention",
        "wall_ms": 598
      }
    },
    {
      "name": "qwen35_sequence_prefill",
      "status": "PASS",
      "duration_us": 555273,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (qwen35_sequence_prefill)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "prefill",
        "wall_ms": 555
      }
    },
    {
      "name": "f16_kv_contiguize",
      "status": "PASS",
      "duration_us": 701159,
      "iterations": 1,
      "parity": {
        "reference_gemv": "CPU reference (f16_kv_contiguize)",
        "logit_cosine_similarity": 0.999999,
        "max_absolute_delta": 0,
        "relative_l2": 0,
        "argmax_exact": false,
        "passed": true
      },
      "metrics": {
        "category": "kv_cache",
        "wall_ms": 701
      }
    }
  ],
  "ablations": [
    {
      "dimension": "target",
      "feature": "cpu_vs_vulkan_gpu",
      "baseline_arm": {
        "name": "cpu_q4_reference",
        "latency_us": 82831,
        "allocated_bytes": 50135040
      },
      "candidate_arm": {
        "name": "vulkan_gpu_q4k",
        "latency_us": 459,
        "dram_bandwidth_gbps": 117.1,
        "allocated_bytes": 50135040
      },
      "speedup": 180.45969498910677,
      "lift_ratio": 180.45969498910677,
      "cosine_parity": 0.9999999999986565,
      "verdict": "VERIFIED_LIFT"
    },
    {
      "dimension": "topology",
      "feature": "fused_vs_discrete_norm_matmul",
      "baseline_arm": {
        "name": "discrete_rmsnorm_then_matmul",
        "latency_us": 34255
      },
      "candidate_arm": {
        "name": "fused_rmsnorm_matmul",
        "latency_us": 21080
      },
      "speedup": 1.625,
      "lift_ratio": 1.625,
      "cosine_parity": 0.999999,
      "verdict": "VERIFIED_LIFT"
    },
    {
      "dimension": "quantization",
      "feature": "quant_q4k_vs_q8_vs_f32",
      "baseline_arm": {
        "name": "f32_dense_weights",
        "latency_us": 1820,
        "allocated_bytes": 356515840
      },
      "candidate_arm": {
        "name": "q4k_super_blocks",
        "latency_us": 428,
        "allocated_bytes": 50135040
      },
      "speedup": 4.252336448598131,
      "lift_ratio": 4.252336448598131,
      "cosine_parity": 0.999998,
      "verdict": "VERIFIED_LIFT"
    },
    {
      "dimension": "quantization",
      "feature": "quant_q2k_vs_q4k",
      "baseline_arm": {
        "name": "q4k_super_blocks",
        "latency_us": 428,
        "allocated_bytes": 50135040
      },
      "candidate_arm": {
        "name": "q2k_super_blocks",
        "latency_us": 265,
        "allocated_bytes": 29245440
      },
      "speedup": 1.6150943396226416,
      "lift_ratio": 1.6150943396226416,
      "cosine_parity": 0.999996,
      "verdict": "VERIFIED_LIFT"
    },
    {
      "dimension": "residency",
      "feature": "device_local_vs_host_visible",
      "baseline_arm": {
        "name": "host_visible_streaming",
        "latency_us": 1420,
        "allocated_bytes": 50135040
      },
      "candidate_arm": {
        "name": "device_local_pool",
        "latency_us": 428,
        "allocated_bytes": 50135040
      },
      "speedup": 3.317757009345794,
      "lift_ratio": 3.317757009345794,
      "cosine_parity": 1,
      "verdict": "VERIFIED_LIFT"
    },
    {
      "dimension": "layout",
      "feature": "strided_vs_contiguized_f16_kv",
      "baseline_arm": {
        "name": "strided_f16_kv_camping",
        "latency_us": 68433,
        "dram_bandwidth_gbps": 28.4,
        "allocated_bytes": 67108864
      },
      "candidate_arm": {
        "name": "contiguized_f16_kv_scratch",
        "latency_us": 25440,
        "dram_bandwidth_gbps": 184.2,
        "allocated_bytes": 134217728
      },
      "speedup": 2.6899764150943395,
      "lift_ratio": 2.6899764150943395,
      "cosine_parity": 1,
      "verdict": "VERIFIED_LIFT"
    }
  ],
  "digest": "sha256:a8a3cdabebd068cf8c971f2a2f218ac4b4de1a671e8ae8a756dc14dc0471f203",
  "verified": true
}
