Does anyone know if Pytorch will support RDNA 2 GPUs? From the documentation, it seems that Pytorch relys on ROCm to run, yet some people have been saying that AMD has abandoned ROCm for RDNA, and is instead focusing on software for their server compute card line up, CDNA.
Also, will Pytorch support DirectML? I’ve read that tensorflow support its, and it allows support for APUs, AMD and Nvidia GPUs. I’ve also it reads its runs faster. I’m not sure if I understand the concepts of ROCm and DirectML correctly, so correct me if my understanding is wrong.
2 Likes
Unfortunally I don’t see Pytorch adopting Directml, with regards to ROCm I’ve been following this for over a year now and people still don’t have support for RDNA 1, maybe this will change with RDNA 2, but I doubt, it Basically Cuda, Rocm and Directml are APIs that provide fast matrix multiplication on a given platform, I like directml because on Windows at least is hardware agnostic, 'cuz it runs on top of Directx, when I heard about Rocm a few years ago I thought that this would finally be the platform to rule them all, but it runs only on Linux and in older hardware, CUDA have been very consistent that’s why I think the PyTorch team will not try to add another runtime for its framework, OpenCL would be the perfect candidate to a heterogeneous computing platform, but it didn’t work out well at the end, plaidml is something that you should look into, unfortunally PyTorch is a CUDA only framework at the moment.
1 Like
There is now support for PyTorch with directml: Introducing PyTorch-DirectML: Train your machine learning models on any GPU - Windows AI Platform