Fastest GPU kernels, written from scratch.

H100 BF16 Matrix Multiplication

Matrix multiplication of square bf16 matrices, accumulated in fp32.

N=4096
Kernel: 763 TFLOPs
cuBLAS: 716 TFLOPs

N=8192
Kernel: 808 TFLOPs
cuBLAS: 795 TFLOPs

Explanation in https://cudaforfun.substack.com/p/outperforming-cublas-on-h100-a-worklog

To run:

make matmul && out/matmul

Example kernels are in h100/matmul/, and orchestration is in h100/matmul.cu.

GB300 NVFP4 Matrix Multiplication

The NVFP4 example contains ten complete kernel snapshots, a shared correctness and benchmark runner, and a cuBLASLt baseline.

make nvfp4
./out/nvfp4 8192 8192 8192 --rounds 5 --cooldown-sec 8

make nvfp4-ladder
./out/nvfp4-r0 8192 8192 8192 --rounds 5 --cooldown-sec 8
./out/nvfp4-r9 8192 8192 8192 --rounds 5 --cooldown-sec 8

This requires a GB300/B300 and CUDA 13.1. See gb300/nvfp4/README.md for every optimization rung, scheduler controls, correctness-only runs, and benchmark modes.

H100 Sum Reduction

We compute sum of 2^30 elements.

To run:

make sum && out/sum

Kernel: 3240.11 GB/s
cub Library: 3193 GB/s

The kernel is in h100/sum.cu.