CUDA Matmul Kernel on B200 Beats cuBLAS by 6%
Author: @paulWilliamChan
Date: 2026-04-27
Summary
Hand-written matmul kernel in pure CUDA/PTX on B200 beats cuBLAS by 6% at M=N=K=8192. Demonstrates that hand-tuned kernels still matter for frontier performance.
See Also
- thunderkittens-gpu-kernels — GPU kernel engineering
- gpu-mode-and-kernel-engineering — GPU programming resources