Comment by mistymountains

Unless you develop in CUDA, you can easily train code (e.g. PyTorch) written for training on Nvidia hardware on AMD hardware. You can even keep the .cuda() calls.