Hi, I've been busy trying to add packages like tensorflow & pytorch and I wanted to add GPU support. I've tried a lot of different options with ROCm but I just get errors in every way I'm trying it. Which is quite a shame as you promote this maker board as a powerful makerboard that can run deep neural networks. Do you know solutions to run tensorflow through gpu?
I have looked into this and tried multiple methods to get the ROCm working like different kernels and such but with no luck. A deep nural network would only consist of two or more hidden layers. Honestly for training your own models on this GPU i think it would be extremely slow. I have a nividia laptop that can run AAA games and have tested out several different models i have trained and even that is pretty slow still. I believe it is a bit misleading on their end, you can run trained models fairly easily, even just on the cpu, like openCV's dnn modules, but for training your own raw model from scratch would be pretty slow.
The board runs OpenCL nicely, on Windows 10 at least, that is because AMD is not enabling the OCL part of their drivers yet for their V1000 chips. I am also waiting eagerly for this to change at some point and it will sooner or later. Then there is this https://bruhnspace.com/en/bruhnspace-rocm-for-amd-apus/
I have a feeling that AMD does not want to fall under the same mistake as Intel did... That is, its inexpensive ATOM line is doing such a good job that it can do a 90% of the main profit maker Core2 families of CPUs. I guess AMD as well as Intel, even the "ARM world" will continue fine-tune their strategy (to add/romove low end embedded APU family of products.)
