For inference (running previously-trained models that need lots of RAM), the desktop could be useful, but I would be surprised if training anything bigger than toy examples on this hardware would make sense because I expect compute performance to be limited.
Does anyone here have practical recent experience with ROCm and how it compares with the far-more-dominant CUDA? I would imagine that compatibility is much better now that most models are using PyTorch and that is supported, but what is the performance compared to a dedicated Nvidia GPU?
For inference (running previously-trained models that need lots of RAM), the desktop could be useful, but I would be surprised if training anything bigger than toy examples on this hardware would make sense because I expect compute performance to be limited.
Does anyone here have practical recent experience with ROCm and how it compares with the far-more-dominant CUDA? I would imagine that compatibility is much better now that most models are using PyTorch and that is supported, but what is the performance compared to a dedicated Nvidia GPU?
ROCM is complete garbage. AMD has an event every year that “Pytorch works now!” and it never does.
ZLUDA is supposedly a good alternative to ROCM but I have not tried it.