News summaryNetEase FinanceNov 14, 2026, 02:31 AM
AMD Shows HRX Loom Accelerating Large-Model Migration
AMD engineers used HRX Loom and an agentic AI workflow to migrate 50 popular Hugging Face models to AMD consumer graphics hardware in three to four days, spanning 18 architectures. [[fact:fact-imrrigciz]]For 32K-token long prompts, HRX Loom ran 1.5 to 2.1 times faster than the Llama.cpp ROCm backend. [[fact:fact-324a1n7ll]]In the same type of test, it showed a 1.9-times performance advantage over the Vulkan backend. [[fact:fact-ftw7lv0sl]]HRX Loom remains an experimental AMD project. [[fact:fact-w5ssouv4y]]The project also supports CPU and NPU execution. [[fact:fact-a94ij7isp]]The experiment results were published on GitHub's Loom-Fun page. [[fact:fact-vopwhd1cz]]