Publish hopper_tensor_core_bf16_perf kernel and performance card

7d10385 verified about 2 months ago

1.47 kB

license: apache-2.0
tags:
  - cuda
  - kernel
  - gpu-optimization
  - hpc

hopper_tensor_core_bf16_perf

Hopper BF16 baseline perf lane at 4096^3 with reference disabled to isolate steady-state Tensor Core throughput.

This repository contains the standalone CUDA source for the hopper_tensor_core_bf16_perf 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

Kernel	GPU / architecture	Shape	Best recorded result	Evidence
`hopper_tensor_core_bf16_perf`	sm90	4096x4096x4096	5.118 ms / 26.855 TFLOPS	Measured on sm90, shape 4096x4096x4096; evidence `hopper-sm90-80-120-async-candidates-20260421T201129Z.json`.

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/kernel.cu.
Original lane tags: cuda, matmul, hopper, sm90, prototype, tensor-core, bf16, perf, next-loop.

Build/run contract

{nvcc} -O3 -std=c++17 -lineinfo -DPYC_HOPPER_TENSOR_CORE_USE_BF16=1 -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