Lemonade 11.9, an open-source local AI server, has been released with experimental support for AMD’s ROCm HRX backend, initially for Radeon RX 7900 series and Strix Halo APUs. This new backend aims to improve performance on AMD hardware by providing a more optimized and client-focused subset of ROCm.
The latest version of Lemonade focuses on providing “100% free and private” AI use through local hardware, including GPUs, CPUs, and NPUs. The inclusion of the Llama.cpp HRX back-end, working with the AMD HRX Runtime, is a key feature of the 11.9 release. HRX is described as a new development within the ROCm ecosystem and part of AMD’s Loom/Hyperloom efforts, announced at the AMD Advancing AI event.
According to the hrx-system GitHub repository, “The HRX System is a collection of minimal runtime components providing an alternative implementation of HIP as it presently ships within ROCm. It provides a common substrate for low latency, high performance integration with AMD’s GPU, NPU, and CPU products.” AMD engineer Stella Laurenzo explained in a GGML discussion that the development of HRX was motivated by the costs of not having AMD native backends and optimized client libraries for optimal performance on their silicon.
Laurenzo also noted that ROCm, originally developed for the datacenter market, often faced challenges when integrating and deploying to client operating systems. HRX is intended to be a lighter, more focused subset of ROCm, optimized for these use cases. It represents a shift towards Loom IR as a custom intermediate representation language for AMD hardware, building on previous work with MLIR and SPIR-V.
Read the original coverage
💬 Comments
📜 Comment Policy