Would it be possible to run AI on an AMD GPU instead of Nvidia?
Yes, it’s about as easy as on Nvidia
Not nearly as flexible though.
I have a 7900XTX I would like to use for AI. I tried to get ollama working on my windows gaming PC this last weekend through Docker and WSL, but that was a pain.
It seems that pytorch might have worked but I still need to try out TensorFlow. Both of those ecosystems seem a hair fractured when it comes to AMD.
While I am sure it can be done, its nowhere near the point and click experience like it was on my laptop with an Nvidia card.
Removed by mod
I’ll keep trying, but I don’t know how far I’ll get.
I am still going to build a dedicated machine for AI work, so I am not sure if I want to break my gaming rig too hard. All of that mess would be so much easier on a Linux rig with none of the windows fluff getting in the way.
That’s the opposite of the feedback I got. AMD claims to support all of the transformers library but many people report this to be a lie.
I am in no love of companies that establish de-facto monopolices, but that is indeed what NVidia has right now. Everything is built over CUDA, AMD has a lot of catch-up to do.
I have the impression that Apple chips support more things than AMD does.
There are some people making things work on AMD, and I cheer to them, but let’s not pretend it is as easy as with Nvidia. Most packages depend on cuda for gpu acceleration.
Things might be about to change: https://www.phoronix.com/review/radeon-cuda-zluda
Can’t wait! But really, this type of things is what makes it hard for me to cheer at AMD:
For reasons unknown to me, AMD decided this year to discontinue funding the effort and not release it as any software product. But the good news was that there was a clause in case of this eventuality: Janik could open-source the work if/when the contract ended.
I wish we had a champion of openness but in that respect AMD just looks like a worse version of NVidia. Hell, even Intel has been a better player!
That’s actually super interesting and potentially game changing! thanks!
I just got DirectML working with torch in WSL2 which was fairly painless.
I am wondering if that isn’t a better option than trying to emulate CUDA directly. Love it or hate it, Microsoft did do a fairly good job wrangling in different platforms with DirectX.
Over my dead body.
This is especially true in the research space (where 90% of this stuff is being built :)
When I ran ollama on my amd GPU PC it said there was no nvidia GPU.
