# Zyphra's open FP8 attention kernels run up to 1.76x faster than AMD's own on the MI355X

> ZYFA covers FP8 and BF16 prefill attention for AMD Instinct MI355X, with 62% to 78% lower error at head dimension 128, Zyphra says.

- URL: https://devisdead.ai/news/zyphra-s-open-fp8-attention-kernels-run-up-to-1-76x-faster-than-amd-s-own-on-the
- Published: 2026-11-08T23:35:42Z
- Author: DevIsDead Newsroom
- Section: Dev
- Tags: amd, zyphra, gpu kernels, open source
- Originally reported by MarkTechPost (https://www.marktechpost.com/2026/11/06/zyphra-zyfa-fp8-flashattention-amd-mi355x/).

Zyphra has published ZYFA, a set of open-source attention kernels for the prefill stage, in both FP8 and BF16, built for AMD's Instinct MI355X, according to [MarkTechPost](https://www.marktechpost.com/2026/11/06/zyphra-zyfa-fp8-flashattention-amd-mi355x/).

On real model activations, Zyphra says its FP8 kernels run 1.16 to 1.76 times faster than AITER, AMD's own kernel library, with 62% to 78% lower error at head dimension 128. Prefill is the part of inference that processes the prompt, so faster attention there shortens time to first token, especially on long inputs.

Details are thin beyond those numbers, and they're Zyphra's own measurements. But the direction matters. Much of Nvidia's lead in AI comes from software: years of hand-tuned kernels that make its hardware fast out of the box. Every open kernel that beats the vendor's own on AMD chips chips away at that moat and makes non-Nvidia hardware a more realistic option for running models.

If you're evaluating MI355X capacity for inference, it's worth benchmarking ZYFA against AITER on your own workloads before settling on numbers.

**[Scoreboard](/scoreboard) update:** AI +1. An open kernel beats AMD's own on AMD's own chip.

Stay undead.