Hello everybody, I hope your Sentinel build plans are coming along!

Today, I wanted to take some time to bring you up to speed on the driver landscape. Because, while Raspberry Pi’s are neat little machines for building a lightweight home server that packs a punch, its Achilles heel seems to be GPU drivers.

AMD has a long history of open source kernel drivers, and it may therefore come as no surprise that they have, by far, the best support for ARM platforms. It only requires a tiny patch to tweak some cache coherencey, and then pretty much any AMD GPU will work on the Raspberry Pi. There is even some work underway to upstream this patch into the official Linux kernel or the Raspberry Pi fork. Long story short, AMD support on the pi is pretty good (but note that ROCm is not supported)!

Intel also has an open source driver, but, being Intel, their ARM support is a bit less mature. There is a patch to disable some x86 bits that make the driver work, but only with 4kb pages and some minor glitches as far as I know. When I last tried to run llama.cpp, it told me "aaaaaaaaaaaaa" so while not as robust as AMD support, it’s also just a minor patch that has been posted to the Linux kernel mailing list.

For NVIDIA, the story is not as simple. The open source Nouveau driver doesn’t seem to work at all, and the proprietary driver doesn’t work either. There is an unofficial fork of NVIDIAs open-gpu-kernel-modules with extensive and complex modifications, but with no official endorsement or clear path to getting upstreamed. I personally haven’t tried them, but as usual Jeff Geerling has, and it seems to work well enough to have some fun with. Though, for the above reasons, the world’s most popular GPU brand is the least recommended for Raspberry Pi in my opinion.

Given all that, my conclusion is that if you have an existing GPU you want to use, all three vendors can be made to work, but if you have a choice, stick with AMD.