{
  "device": "NVIDIA GB10",
  "compute_capability": "SM121",
  "cuda": "13.0.88",
  "comfy_kitchen": "0.2.31",
  "cutlass_dsl": "4.6.2",
  "existing_interface": {
    "registered_schema": "scaled_mm_nvfp4(a, b, tensor_scale_a, tensor_scale_b, block_scale_a, block_scale_b, bias, output_dtype_code, alpha) -> Tensor",
    "backend": "cublas_gemm_blockwise_fp4",
    "allocates_complete_activation_qdata": true,
    "allocates_complete_activation_block_scales": true,
    "allocates_complete_bf16_output": true,
    "accepts_incremental_tiles": false,
    "accepts_user_output": false,
    "accepts_custom_epilogue": false
  },
  "owned_kernel_support": {
    "sm121_blockscaled_nvfp4": true,
    "persistent_cooperative_schedule": true,
    "pipeline_primitives": true,
    "custom_software_a_producer": "requires editing the CuTe DSL mainloop",
    "cluster_shape": [1, 1, 1],
    "activation_tile_multicast": false
  },
  "stock_cute_baseline": {
    "shape_mnkl": [1024, 1024, 1024, 1],
    "tile_shape_mnk": [128, 128, 128],
    "a_dtype": "Float4E2M1FN",
    "b_dtype": "Float4E2M1FN",
    "scale_dtype": "Float8E4M3FN",
    "scale_vector_size": 16,
    "accumulator_dtype": "Float32",
    "output_dtype": "BFloat16",
    "execution_microseconds": 16.966399550437927,
    "tflops": 126.57273817087317,
    "reference_check": "PASS"
  },
  "decision": "Use a two-phase exact design: global BF16 scale reduction followed by an owned CuTe DSL GEMM with a software A/SFA producer. Measure N-tile reuse before expanding beyond one fixed tile.",
  "design_document": "NVFP4_STREAMING_DESIGN.md"
}