LaunchDevelopersNovember 6, 2026

Zyphra releases ZYFA FP8 FlashAttention kernels for AMD MI355X

Read original source →marktechpost.com

Open-source (Apache 2.0) FP8/BF16 prefill attention kernels run 1.16-1.76x faster than AMD's AITER on real model activations, with 62-78% lower FP8 error at head dim 128. Peak hit 2,672 TFLOP/s FP8 at 32K tokens, 53% of datasheet peak versus AITER's 39%.

1 source

More stories today

Open the live feed