Intel Mac Pro, Vega II revive local AI

Hot take: the machine is usable, but macOS is the wrong place to expect modern local-AI performance from it. The hardware is interesting; the software stack is what makes it awkward.

• –LM Studio officially excludes Intel Macs, and Ollama’s macOS path is Apple Silicon or CPU-only x86, so staying on macOS likely leaves the Vega II underused.
• –Linux is the practical route if you want GPU acceleration: Ollama documents AMD support on Linux and Windows, including Vega II, and llama.cpp supports Vulkan and SYCL backends.
• –AMD’s own specs put Vega II at 32GB HBM2 and up to 1 TB/s bandwidth, which is enough to host sizable quantized models even if raw speed trails modern cards.
• –Community reports around MI50-class Vega 20 parts suggest decode speeds in the single digits to low teens tok/s for many real-world setups, with prompt processing usually much faster than generation.
• –For MCP servers, this is best treated as a dedicated LAN inference appliance with a local OpenAI-compatible endpoint, not as a fast interactive workstation.