AMD Leverages AI + HRX Loom In Experiment For Quickly Porting 50 LLMs To Client GPUs

Following the recent release of HRX Loom v0.1 and leveraging an AI agent workflow, a team at AMD experimented with bringing 50 trending Hugging Face models over to AMD client GPUs. They accomplished this feat in just around three days, thanks to their HRX software work paired with the power of Ryzen AI Max "Strix Halo" hardware. AMD engineers Jeremy Fowers and Ben Vanik led an experiment to bring 50 popular Hugging Face models across 18 architectures over to AMD client GPUs in just three to four days. This experiment was done to show the potential of bringing new as well as niche models to AMD hardware faster thanks to HRX Loom and agentic AI. They also found that there was a medium 1.5~2.1x speed-up for 32K token prompt processing compared to Llama.cpp's ROCm back-end. Compared to Llama.cpp's Vulkan back-end was a 1.9x speed-up. HRX Loom remains an experimental project for now at AMD but showing great potential given these results. HRX Loom is not only for GPUs but also supports CPU and NPU execution too. Those curious about this AMD HRX Loom experiment can find their results via the Loom-Fun GitHub page. As part of this work is the loom-skills deliverable for equipping Claude Code, Codex, or Pi agent with the necessary bits to author and serve optimized model programs to AMD client GPUs -- with the current focus/testing being on the wonderful Strix Halo hardware with the likes of the Ryzen AI Halo or Framework Desktop. More background information on AMD's Loom can be found via the documentation.