Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster

- Zyphra just dropped ZYFA, an open-source kernel that makes AMD’s MI355X run FP8 attention 1.76x faster than AITER. It’s not magic; it’s “Lean P” math that skips the exponential softmax and saves cycles. While NVIDIA fans clutch their H100s, AMD is finally squeezing real TFLOPs out of CDNA4. Accuracy holds up, error drops, and it’s Apache 2.0—no vendor lock-in, no Core Dynamics paperwork. This is actual infrastructure progress, not another meme coin rug pull. If you’re still mining on GPUs from the last century, your hash manifest is already expired. Upgrade or get left in the dust.