AMD ports 50 Hugging Face LLMs to client GPUs in three days using HRX Loom and AI agents

AMD engineers Jeremy Fowers and Ben Vanik conducted an experiment demonstrating that 50 popular Hugging Face large language models spanning 18 architectures could be ported to AMD client GPUs in approximately three to four days. The work relied on HRX Loom v0.1, an experimental AMD software project released shortly before this effort, paired with an AI agent workflow and Ryzen AI Max "Strix Halo" hardware for execution and validation.

The experiment produced a deliverable called loom-skills, which equips AI coding agents including Claude Code, Codex, and Pi with the capability to author and serve optimized model programs targeting AMD client GPUs. Performance testing against established inference backends showed HRX Loom achieving a 1.5 to 2.1x speed-up for 32K token prompt processing compared to Llama.cpp's ROCm back-end, and a 1.9x improvement over Llama.cpp's Vulkan back-end. HRX Loom supports execution across GPU, CPU, and NPU targets, with current development focus on Strix Halo hardware including the Ryzen AI Halo and Framework Desktop platforms.

HRX Loom architecture and performance results

• HRX Loom v0.1 scope: The experimental framework supports execution on GPU, CPU, and NPU, not limited to discrete or integrated graphics alone, enabling flexible deployment across AMD's heterogeneous compute stack.
• AI agent integration: The loom-skills deliverable provides tooling for Claude Code, Codex, and Pi agents to automatically generate and serve optimized model programs, reducing manual porting effort for new and niche models.
• Performance against Llama.cpp: For 32K token prompt processing, HRX Loom measured a 1.5 to 2.1x speed-up over Llama.cpp's ROCm back-end and a 1.9x speed-up over its Vulkan back-end, indicating competitive or superior kernel efficiency on AMD hardware.
• Target hardware: Development and testing focused on Ryzen AI Max "Strix Halo" silicon, specifically the Ryzen AI Halo and Framework Desktop systems, AMD's current flagship client APU with substantial integrated graphics and NPU resources.
• Model coverage: The experiment encompassed 50 trending Hugging Face models across 18 distinct architectures, suggesting HRX Loom's abstraction layer generalizes across transformer variants without per-model manual tuning.

Why it matters

For Linux users running AMD hardware, HRX Loom represents a potential acceleration path for local LLM inference that outperforms the widely-used Llama.cpp on ROCm and Vulkan paths, reducing the historical gap between AMD and NVIDIA CUDA ecosystems for AI workloads. If HRX Loom matures from experimental status, it could lower the barrier for developers and distributions to ship optimized AI tooling on AMD client GPUs, APUs, and NPUs without waiting for upstream framework support or manual kernel optimization.