I don't know if you really need the full ROCm stack just to throw matrices at a card. Inference requires less "scaffolding", but maybe Fujitsu just want to outs

I don't know if you really need the full ROCm stack just to throw matrices at a card. Inference requires less "scaffolding", but maybe Fujitsu just want to outsource the whole stack and forget about it. My only worry would be that the software model of compute that ROCm has is going to drastically different than the hardware model of compute the Fujitsu chip has. This would be like using one of those CUDA compilers that target AMD hardware. Sure you can produce something that will run, but depending on what you are compiling, it will run dog slow.