I don't know about AI performance (does this happen only because of the overhead of providing CUDA through rocm+HIP?), but I was just checking and at least in m

I don't know about AI performance (does this happen only because of the overhead of providing CUDA through rocm+HIP?), but I was just checking and at least in my country (Brazil), for any given memory size (12GB, 8GB, 4GB) I can find cheaper AMD GPUs than NVidia GPUs Here I'm considering that the main constraint is VRAM and while stable diffusion now runs even on GPUs with 2GB RAM, there's always new developments that require more VRAM (for example, Dreambooth requires 12GB as of today)