Currently using Aurora Linux, would be willing to switch to anything but the ubuntu, vanilla fedora, bazzite are pretty much behaving the same way.
I am trying to use multiple GPUs, the iGPU and an eGPU. the eGPU through USB4 (although I have a riser cable to plug it directly and a way to provide external power). Drivers are up i can se the card on the system The issue that is driving me nuts is that the nvidia card is showing up as GPU0 always; regardless of any kernel parameter, hard coding paths, udev rules, that I could think to change so that it is listed as GPU1.
My issue is that I dont want to add a launch parameter to every app. I want ALL of the system to use the iGPU except for 1 container.
I don’t have an answer to your problem, but I would be interested in hearing how you’re getting on with your eGPU. Are you using it for graphics or AI compute?
I use it for compute. Mostly playing around with machine learning.
I still have not solved it. Seems in general multiple GPUs in linux is still a mess. Seems KDE and wayland are working on it.
I read here in the framework community someone used the GPU pcie address instead of the GPU position in the env variables to achieve a similar behavior I have not tried it yet.
Even through USB4 and addings several seconds of model loading some tasks just run much faster on nvidia. Mind you, im not using a ‘good’ GPU, it is an RTX 3060 12GB.
So currently I run most things on the iGPU except for those rare instances where the Nvidia does give me significantly better times and does not require more than 12GB of vram.
Once I test a few more things I will post what did work and what didn’t.
I’ll be interested to see what you find. I think I’ll order a FW Desktop in a few weeks. After a fair bit of research, I am expecting the unified arch to produce better reasoning that than an eGPU, just on the basis that more advanced models will fit in 60-80GB of VRAM. I think your card will run much faster, but at the cost of only being able to fit a smaller model; you’ll be trading reasoning for speed.
But if you can find an inference task that is parallelisable, of course you could run both GPUs at the same time.
Adding a note to this thread: I had not appreciated previously that the USB4 interface should outperform the PCIe interface by a significant margin. Thus, I am hoping there are no issues in getting the two GPUs working happily in Linux at the same time.
Hopefully I will get to try it myself, though I want to be experimenting with AI, not fighting driver-level issues
