RSSAmplifier

Blog

ML Perf Notes

/RSS feed ↗4 posts

Latest posts

MXFP8 GEMM: Up to 99% of cuBLAS performance using CUDA + PTX

I recently did a deep-dive on writing GEMM kernels with just CUDA + PTX for Ampere, Hopper, and Blackwell GPUs, culminating in a MXFP8 GEMM kernel which achieves up to 99% of cuBLAS (torch._scaled_mm) depending on the problem shape - see microbenchmarks below, measured with: B200 GPU, 1000W power CUDA 13.0 PyTorch version: 2.11.0+cu130 K-major input operands, with float32 output

Debugging deadlocks in warp-specialized GEMM kernels with CUDA-GDB

While writing warp specialized GEMMs for Blackwell with CUDA + PTX, I’ve noticed there are some great resources on efficient kernel designs for this architecture, but a lack of resources on debugging the often cryptic or even completely opaque CUDA errors that arise during the kernel development process.

An illustrated deep-dive into how the compute and comms in TP+SP are overlapped using Async TP

<script type=”text/javascript”

An illustrated deep-dive into Megatron-style tensor parallelism

<script type=”text/javascript”