Zyphra releases open source ZYFA attention kernels for AMD MI355X
Zyphra has released ZYFA, open-source prefill attention kernels for AMD Instinct MI355X GPUs that outperform AMD default kernel library on real model activations.
FTMQ SI, written by our newsroom0 views
Zyphra has released ZYFA, open-source FP8 and BF16 prefill attention kernels for AMD Instinct MI355X GPUs. MarkTechPost reported that the software functions as an attention engine designed for long-context prefill operations. The code has been made publicly available under the Apache 2.0 license. [1]
According to MarkTechPost, benchmarks on real model activations show that ZYFA FP8 kernels run 1.16 to 1.76 times faster than AITER, which is the default kernel library provided by AMD. The report added that at a head dimension of 128, ZYFA reduces FP8 error by 62% to 78% compared to AITER. Zyphra distributed the software alongside a technical paper and implemented it in the attention engine of Zyphra Optimi. [1]
AMD Instinct is a brand of data center GPUs designed to accelerate deep learning, artificial neural networks, and high-performance computing applications. It replaced the FirePro S brand in 2016 as AMD dedicated server hardware line. [3]
In short
- Zyphra released ZYFA as open-source FP8 and BF16 prefill attention kernels under the Apache 2.0 license.
- ZYFA FP8 kernels run 1.16 to 1.76 times faster than AMD default AITER library on real model activations.
- At head dimension 128, ZYFA reduces FP8 error by 62% to 78% compared to AITER.
Sources
Every paragraph above points to the numbered items it rests on. Read the originals here.
- [1]Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x FasterMarkTechPost, 11d ago (the report this story comes from)
Background
- [2]Xiaomi MiMo on Wikipedia
- [3]AMD Instinct on Wikipedia
Our newsroom writes these reports with the help of software, from the 3 sources listed and nothing else, and checks them against those sources. Facts can still be wrong or move on; the originals are the record. Spotted a mistake? Write to daniel@monsterkong.com.
Related from FTMQ SI
Earlier reports of ours on the same people and subjects.
- AI scanning tool uncovers dozens of Linux kernel bugsModels & Products, 2d ago
- Alibaba releases Qwen4-Max model with native vision MoE architectureModels & Products, 1d ago
- Linus Torvalds Shares Views on AI Usage in Open Source ProgrammingModels & Products, 39d ago
More in Models & Products
- Microsoft unveils Copilot Workspaces and ROI tools at Ignite 20265h ago
- Meta deploys LLM test system to catch software bugs in production1d ago
- Alibaba releases Qwen4-Max model with native vision MoE architecture1d ago
- NiCE launches next generation Cognigy platform for agentic contact center automation2d ago
- AI scanning tool uncovers dozens of Linux kernel bugs2d ago
- Cmpnd releases dspy-wsp for weight space programming in DSPy2d ago
Get the day in one email
Reports like this one, the top news of the last 24 hours, every morning. Free, one email a day; readers can comment under every report.
By signing up you agree to our terms and privacy policy. Unsubscribe any time.

Comments
Loading
Join the conversation
Comments are open to readers of our daily email: the top news of the last 24 hours, every morning, free. Sign up and the comment box opens.
Already on the list? Enter the same address and we will send a sign-in link.
By signing up you agree to our terms and privacy policy. Unsubscribe any time.