ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4

Zyphra has released ZYFA, an open-source library providing FP8 and BF16 prefill attention kernels optimized for the AMD Instinct MI355X hardware. These kernels improve computational efficiency by running up to 1.76 times faster than AMD’s existing AITER implementation while simultaneously reducing mathematical error rates. This development provides a performance alternative for accelerating attention mechanisms in large models, which are often limited by compute capacity during long-context processing.