Running Local LLMs (“AI”) on Old AMD GPUs and Laptop iGPUs (Arch Linux Guide)

A straightforward guide on how to compile llama.cpp with Vulkan support on Arch Linux (and Arch-based distros like CachyOS, EndeavourOS, etc). This lets you run models on old, officially unsupported AMD cards and Intel iGPUs.

The same steps work on Debian/Ubuntu, but the package names are different.

Here’s how I’m running models on 3 × AMD Radeon RX 580 8 GB (24 GB VRAM total) without ROCm in 2025.