Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster
Zyphra has released ZYFA (Zyphra FlashAttention on AMD), a set of FP8 and BF16 prefill attention kernels for AMD Instinct MI355X GPUs. ZYFA is Zyphra’s FP8 attention engine for long-context prefill, and the code is open source under Apache 2.0. On real model activations, its FP8 kernels run 1.16-1.76x faster than AITER, AMD’s default kernel library. At head dimension 128, ZYFA’s FP8 error is 62-78% lower than AITER’s. TL;DR Framework: It is a kernel library. Causa
Get smart on it
ZYFA is a kernel library that speeds up a specific part of AI model computation called attention on AMD graphics processors. The library uses a lower-precision numerical format called FP8 instead of the standard format, which allows it to run certain calculations up to 1.76 times faster than AMD's default library while reducing computational errors by 62 to 78 percent. The performance improvement matters because attention is a bottleneck in processing long sequences of data, though real-world gains vary depending on the specific model being used. ZYFA is released as open-source code and is designed specifically for AMD's MI355X GPU, running only the initial "prefill" stage of attention rather than the full computation pipeline.
Relevant courses
Want to go deeper than the news? Explore live, cohort-based AI courses taught by practitioners.
Browse AI courses on MavenRelated news
We Visited China’s Stargate in Inner Mongolia
Stargate Abilene has become the symbol of America’s AI buildout. It is a gigawatt-scale campus on the Texas plains built for OpenAI. China has campuses on a similar scale, though few people outside the country have had a look inside. In September, we visited one, and we found the speed, scale, ultra low cost, and deep vertical integration astonishing.During the trip, we toured Inner Mongolia, China’s fastest-growing AI datacenter region. It's the “West Texas” of China, bu
Elon Is Doing It Again: SpaceX Will Be The World’s Biggest AI Cloud
We wrote at length about how SpaceX’s plans for 10GW in 2027 was feasible, paving the way for +300b+ ARR. We received a lot of pushback from folks calling it impossible. Now, just two months later, we’re already seeing 3.6GW under construction across multiple sites, both existing and new. SpaceX has even revised its estimate up to ~13GW in 2027. This would be the largest amount of datacenter capacity ever built by any one company in that timeframe, roughly equivalent to all of Google
Data centers lost the midterms
Data center developers are poised to face a more restrictive building environment after voters across the country voted in candidates who favor moratoriums and ballot measures that put more obstacles on their expansion, delivering a stinging blow to a central pillar of the AI industry. Of voters polled by NBC News on Election Day, 76 percent said they opposed data centers in their communities. Dozens of local communities voted directly to restrict development through ballot mea
Also reported by Wired AI and TechCrunch
Crusoe sells troubled Wyoming data center project to Google
The deal highlights the challenges of developing massive AI data centers.
What Huawei’s Horinger data center reveals about China’s compute gap
Huawei’s AI data center in Horinger, Inner Mongolia, uses Ascend chips and has an estimated AI computing capacity that is just 7% that of the world’s largest AI data center, SpaceX’s Colossus 2 in Tennessee.
Protesters break into Microsoft data center site in Amsterdam
Members of the action group Geef Tegengas put a banner on a crane during a protest. | Image: Getty Images Protesters have broken into a construction site of a new Microsoft data center in Amsterdam this morning. Members of the Geef Tegengas group helped organize the protest against the construction of three 85-meter data center towers in the outskirts of Amsterdam. Reuters reports that around 100 people made it into the construction site, with some scaling a crane to place a "fu