hopper_tensor_core_bf16_async_wide_k64

Hopper BF16 async WMMA lane keeping the 64x128 CTA but doubling stage depth to 64 so each stage amortizes more feed and loop overhead.

This repository contains the standalone CUDA source for the hopper_tensor_core_bf16_async_wide_k64 lane from the PyC kernel lab. It is a source artifact for inspection and benchmarking; it is not a precompiled binary and the result below is not a universal ranking.

Performance

The result is reported with the original campaign's timing and correctness context. Compare kernels only when GPU, CUDA version, matrix shape, warmup, repeats, and reference/correctness mode match.

Source

• kernel.cu — copied from kernels/prototypes/hopper/tensor_core_async/kernel.cu.
• Original lane tags: cuda, matmul, hopper, sm90, prototype, tensor-core, bf16, async, cpasync, double-buffered, wide, k64, gap-close.

Build/run contract

{nvcc} -O3 -std=c++17 -lineinfo -DPYC_HOPPER_TENSOR_CORE_USE_BF16=1 -DPYC_HOPPER_TC_WARP_COL_GROUPS=4 -DPYC_HOPPER_TC_TILE_K=64 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90 {source} -o {build_dir}/{name}
{build_dir}/{name} 4096 4096 4096 3 30 1