>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 tal

>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.