My, how the times have changed.
Ive heard this, but Im not convinced. Don't they have more memory, so while the processing is slower there is less fetching of memory outside of the card?
Of course this would require custom compute kernels for same throughput like matrix multiplication, but still.
Where it fails is things that use the compute functions (not vulkan) like blender and many neural network tools. Basically AMDs version of cuda called ROCm. They only support ROCm on a few of their cards while nvidia has cuda on everything on day one.
It's supposed to be out, and still isn't. ¯\_(ツ)_/¯