If you try installing ROCm on a new AMD CPU server, e.g. Genoa, you may note that the driver fails to load on boot up.
In the dmesg you may see something like amdgpu: .... invalid argument
Turning on the iommu worked for us, as described here:
https://community.amd.com/t5/knowledge-base/iommu-advisory-for-amd-instinct/ta-p/484601
