Comment by 65a

You can run inference and training on consumer AMD cards today. It works fine, including llama.cpp, stable diffusion, hugging face transformers, etc. Way cheaper for a given performance/VRAM target as well.