Show HN: Where a hand-written GPU kernel beats the library, and where it can't

By vishal-padia · 2026-07-30 · 1 points · 0 comments

https://kyrieblunders.bearblog.dev/the-kernels-that-were-there/

I've been building streaming inference kernels for Zonos, the open SSM-based (Mamba2) TTS model, on an A10G. This post is two hand-written CuTeDSL kernels in two regimes, and the contrast is the point. The first is a skinny bf16 GEMV for the batch-1 decode projections. It beats…

Open the full discussion on BetterNews