{
  "recorded_date": "2026-05-18",
  "platform": "RTX 3070 Ti Laptop (GA104, sm_86, Ampere)",
  "note": "Local production-library reference baselines measured on this exact machine. Only locally available reference stacks are recorded here; unsupported kernels are marked unavailable instead of estimated.",
  "default_valid_when": {
    "require_no_throttle": true,
    "allow_throttle": ["GpuIdle"],
    "comment": "Same fair-run policy as data/baselines.json."
  },
  "schema": {
    "exe": "(optional) path to reference bench executable",
    "library": "(optional) local reference stack used by this bench, e.g. cuBLAS",
    "match": "(optional) substring of the bench-output line that holds this reference value",
    "project_kernel": "(required for comparisons) path to the project-kernel baseline entry in data/baselines.json",
    "project_config": "(required for comparisons) config key in data/baselines.json to compare against",
    "label": "(optional) human-readable row label for compare_reference.R output"
    ,"peak": "(optional) hardware peak for this workload's precision class, in the same unit as the stored metric"
  },
  "unsupported": [
    {
      "label": "Flash Attention seq=1024 b=8 h=8 d=64",
      "library": "cuDNN SDPA",
      "reason": "Installed cuDNN headers on this machine do not expose the graph-based SDPA frontend used for a direct local attention reference, so this row remains unmeasured locally."
    },
    {
      "label": "GroupNorm SD 320ch",
      "library": "cuDNN",
      "reason": "No direct local cuDNN GroupNorm harness is implemented yet, so this row remains unmeasured locally."
    }
  ],
  "kernels": {
    "kernels/reference/cublas_hgemm/bench.cu": {
      "exe": "kernels/reference/cublas_hgemm/bench",
      "library": "cuBLAS",
      "2048_2048_2048": {
        "ms": 0.600,
        "gflops": 28631,
        "match": "cublas_hgemm",
        "peak": 174000,
        "project_kernel": "kernels/gemm/hgemm/hgemm_16warp.cu",
        "project_config": "2048_2048_2048",
        "label": "HGEMM 16-warp 2048^3"
      },
      "4096_4096_4096": {
        "ms": 4.626,
        "gflops": 29708,
        "match": "cublas_hgemm",
        "peak": 174000,
        "project_kernel": "kernels/gemm/hgemm/hgemm_16warp.cu",
        "project_config": "4096_4096_4096",
        "label": "HGEMM 16-warp 4096^3"
      }
    },
    "kernels/reference/cublas_igemm/bench.cu": {
      "exe": "kernels/reference/cublas_igemm/bench",
      "library": "cuBLAS",
      "4096_4096_4096": {
        "ms": 4.668,
        "tops": 29.44,
        "match": "cublas_igemm",
        "peak": 348,
        "project_kernel": "kernels/gemm/igemm/igemm_pipelined_cpasync.cu",
        "project_config": "4096_4096_4096",
        "label": "IGEMM pipelined cp.async 4096^3"
      }
    },
    "kernels/reference/cusparselt_igemm_sparse/bench.cu": {
      "exe": "kernels/reference/cusparselt_igemm_sparse/bench",
      "library": "cuSPARSELt",
      "2048_2048_2048": {
        "ms": 0.138,
        "tops": 124.28,
        "match": "cusparselt_igemm_sparse",
        "peak": 348,
        "project_kernel": "kernels/gemm/igemm/igemm_sparse_tiled.cu",
        "project_config": "2048_2048_2048",
        "label": "Sparse IGEMM tiled 2048^3"
      },
      "4096_4096_4096": {
        "ms": 0.808,
        "tops": 170.11,
        "match": "cusparselt_igemm_sparse",
        "peak": 348,
        "project_kernel": "kernels/gemm/igemm/igemm_sparse_tiled.cu",
        "project_config": "4096_4096_4096",
        "label": "Sparse IGEMM tiled 4096^3"
      }
    },
    "kernels/reference/cudnn_conv2d/bench.cu": {
      "exe": "kernels/reference/cudnn_conv2d/bench",
      "library": "cuDNN",
      "1_64_64_320_320": {
        "ms": 0.446,
        "gflops": 16910,
        "match": "cudnn_conv2d",
        "tolerance": 0.20,
        "peak": 174000,
        "project_kernel": "kernels/convolution/conv2d/conv2d_implicit_gemm.cu",
        "project_config": "1_64_64_320_320",
        "label": "Conv2d implicit GEMM 1x64x64x320x320",
        "note": "Median of repeated local cuDNN runs clusters around ~16.9 TFLOPS, but first invocation from an idle laptop state can dip into the mid-14 TFLOPS range; use a wider tolerance to avoid false regressions."
      }
    }
  }
}