Strix Halo Guide: AMD Ryzen AI MAX+ 395 Local LLM Setup & Benchmarks
An independent, evidence-backed guide to running large language models locally on AMD Strix Halo (Ryzen AI MAX+ 395) hardware, covering BIOS setup, Ollama, llama.cpp, ROCm and cross-vendor benchmarks.
This is a practical, independently maintained guide (not an official AMD or OEM publication) for setting up and benchmarking large-model inference on AMD Strix Halo systems with a Radeon 8060S GPU and 96GB or 128GB of unified memory. It covers BIOS configuration, Ubuntu 24.04 and kernel setup, Ollama, llama.cpp over Vulkan/RADV, ROCm/HIP experiments, vLLM notes, and capacity evidence for large models up to 70B, 120B and a 284B GGUF.
The repository states its measured claims are backed by raw logs and reproducibility checks, and separates its own measured numbers, for example a measured 20.42 tokens-per-second figure for one model configuration, from unverified community-reported figures. It documents the maintainer's own upstream contributions to related open-source projects (llama.cpp, the Lemonade local-AI server, llmfit and others) as a credibility reference, alongside a companion project website.