[imtringued]

>Explicitly copying data to/from GPUs is a definite nuisance.

CXL allows fine grained shared memory, but people look at the shiny high bandwidth NVLink and talk about how much better it is for... AI.