I have not tried this in a while but no. It requires a very involved setup compared to it's CUDA counterpart and certain functionalities don't work / performance characteristics for specific operations can vary wildly. If it look at the amount of time necessary to deal with it the calculus quickly swings in favor of just purchasing an Nvidia card.

Edit: And some sibling comments have pointed out that it doesn't function on current gen cards.

Nvidia still firmly leads the software game in this regard. I'm really hopeful this will change in the future but I had the same criticisms four years ago when I bought my last AMD card and nothing has changed yet.

https://github.com/RadeonOpenCompute/ROCm#Hardware-and-Softw... Note that ROCm does not have support for even the current-generation of gaming GPUs (Navi).

ROCm is only available on Linux and doesn’t support NAVI GPUs.

Overall if you need to have a solution that just works with any ML framework out there Radeon GPUs aren’t fit for purpose.

Let us know how that works out.