Zyphra Releases Faster Attention Tool for AMD AI Chips
Zyphra has released ZYFA, a free open-source tool that speeds up a key step in AI models on AMD's MI355X chip. In tests on real model data, its FP8 versions ran 1.16 to 1.76 times faster than AMD's own tool, with lower error rates. Developers running large AI models could benefit from the speed gains.
Quick glossaryZyphra, FlashAttention, FP8
- Zyphra
- Zyphra is an AI company that released this open-source tool for AMD chips, as the story describes.
- FlashAttention
- A fast method that helps AI models process long pieces of text using less memory and time.
- FP8
- A compact number format that stores numbers in fewer bits, so AI chips can run faster.
Burning Potato score: 64%