Hi,
Currently a undergrad in computer science but I am looking to build a PC for deep learning. Considering vega cards are coming back to sane pricing now. I am confused between choosing a gtx 107o or a vega 56 with primary purpose being deep learning. CUDA on the nvidia side works great but as all things team green its closed source, where as RCOm from AMD seems to look very promising. Thoughts from anyone with any relative experience using it for tensorflow/pytorch with vega or RCOm in general would be nice, Thanks!
1 Like
OpenCL is not currently supported by Tensorflow nor Caffe nor pyTorch nor Microsoft’s cognitive Toolkit. And Theano, which apparently did support OpenCL has been deprecated because it is no longer being developed. There are efforts to support AMD but they aren’t ready for prime time yet.
AMD’s open source efforts are great and all and I don’t like nvidia myself, but if you are serious about neural networks there is no choice right now.
3 Likes
How soon do you need it? ROCm is interesting, but I’m still waiting to see how it turns out. I think if you need to use it right now, nvidia/CUDA is the least hassle. Unless you want to be a trailblazer/debugger/etc.
See the table with the frameworks supported: https://rocm.github.io/dl.html Though the 2016 date at the bottom of the page isn’t confidence-inspiring. I can never remember where the “official latest” ROCm stuff is supposed to be. I think this was it: https://gpuopen.com/professional-compute/
1 Like
You will see a lot of things ending around that time as that is when they were talking about merging OpenCL into Vulkan. People around then were litterally telling everyone to stop using OpenGL and OpenCL and too switch to Vulkan.
Even though the 2016 part isn’t confidence inspiring the repos are still updated and being worked on also considering amd’s efforts working on the open source drivers is making me seriously consider it. And also HIPtensorflow seems promising. I currently have a 1070ti but I want to get my hands in AMD’s cookie jar as well.
Yes, they do seem to be putting some effort into ROCm and I really hope it succeeds. Problem is almost nobody knows about it, and the information is not well organized/publicized.
Getting a few major frameworks, probably starting with tensorflow, well supported would go a long way.
If you do please report back. I’d like to switch away from nvidia as well.
1 Like
Did some further research about setting up packages for the current price of the vega 56 it doesn’t seem to be worth the hassle right now. I’ll wait for the crypto mining market to crash again and see if I can buy a used card and play around with that.
How literate in terms of Computer Science does one have to be to get those packages running with a Vega 56? I have 3 of them kinda idleing around and I’m kinda curios…
…and if I can add at least something to this project, then I kinda want to.
In terms of getting them to work I am not entirely sure since I haven’t played around with them. In terms of deep learning with cuda its just a simple metapackage on anaconda with an nvidia card. I think with RCOM we’d have manually compile each framework to make it worth with HIP. The documentation is not very detailed as the toolkit is still very new.
And you have 3 vega cards lying around woah
1 Like
I’m on vacation starting Sunday so I can install all the bells and whistles over the weekend but next weekend when I’m back I can take a look at it we can investigate further.
Yeah they were meant for mining but the location has no power nor web access so I have 3 v56 and lying around until they can get used when ever that will be
I would not start with Tensorflow. Instead I would have a look at fastai’s framework which is built on top of pytorch. Tensorflow is only popular because Google has more marketing money than other open source projects.
It took me a long time to try the fast.ai course since I was determined to start from the bottom up so instead of getting my feet wet, I just ended up learning a lot of half baked ideas without understanding how they work practically.
1 Like