We're trying to run some @PyTorch benchmarks for our open-source inference server NOS (https://github.com/autonomi-ai/nos) and wanted to get a deeper understanding of the state of ROCm compatibility vs. CUDA.
I love this because I predicted this question nearly a year ago now. The answer is that the options today are limited. AMD has previously only focused on deploying these cards to HPC clusters. That said, this is changing quickly.
From what I have read, Azure has some MI300x available on their site.
Lamini (US) has some MI210's and MI250's and one box of MI300x.
ElioVP (Europe) might have some MI2X0's and has one box of MI300x.
TensorWave (US) say they have MI300x, but it isn't confirmed in public.
My startup, Hot Aisle (US), will also have a box of MI300x soon (delayed by Chinese New Year), with significantly more on the way in March/April.