Hipfire hits 1,200 tok/s on AMD Strix Halo

// 45d agoOPENSOURCE RELEASE

Hipfire hits 1,200 tok/s on AMD Strix Halo

Hipfire, a Rust-native inference engine for AMD hardware, introduced an experimental MMQ path that boosts prefill speeds by over 3x on RDNA3 GPUs. Benchmarks on Strix Halo systems show throughput jumping to ~1,260 tok/s, matching performance of specialized implementations like llama.cpp.

// ANALYSIS

AMD APU users are seeing significant performance gains without waiting for official ROCm support. This update achieves up to a 3.8x speedup in prefill throughput by targeting the RDNA3/3.5 instruction set with i8 WMMA and tiled matrix-matrix kernels on Strix Halo hardware. The opt-in HIPFIRE_MMQ=1 toggle ensures stability while providing validated acceleration across multiple KV-cache modes.

// TAGS

llmamdrdna3inferencerustgpubenchmarkquantizationstrix-haloopen-source

DISCOVERED

45d ago

2026-04-28

PUBLISHED

45d ago

2026-04-28

RELEVANCE

8/ 10

AUTHOR

Own_Suspect5343

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

VIDEO12m ago

Gemma Showcases Reachy Mini Running Gemma 4

Google Gemma shared a demonstration of Reachy Mini, an open-source desktop robot developed by Pollen Robotics and Hugging Face, showcasing real-time voice conversation powered by Gemini Live. The demo highlights the robot's physical responsiveness and concludes with a preview of it running entirely locally on the upcoming Gemma 4 model.

MODEL1h ago

ByteDance's Seedance 2.0 multimodal video generation model showcases highly expressive AI-generated human emotions.

Seedance 2.0 is an advanced, multimodal AI video generation model developed by ByteDance that has gained significant attention in the creator community for its realistic portrayal of human emotions. Unlike older generation pipelines, Seedance 2.0 allows creators to combine text, image, video, and audio inputs in a single unified architecture. The model is capable of outputting synchronized video and audio with precise narrative control, allowing creators to prompt specific emotional intensities (e.g., joy, sadness, hesitation) to achieve highly nuanced facial expressions and body language in generated characters.

OPEN SOURCE1h ago

Ponytail cuts AI agent over-engineering

Ponytail is an open-source skill and configuration tool for Claude Code, Cursor, Cline, and other AI agents designed to promote a "lazy senior developer" mindset. By establishing a strict hierarchy of checks—prioritizing native platform features, standard libraries, and existing dependencies—it prevents AI tools from generating unnecessary boilerplate, resulting in cleaner codebases, lower token costs, and faster output generation.

Hipfire hits 1,200 tok/s on AMD Strix Halo