Comment on Porting HPC Applications to AMD Instinct MI300A Using Unified Memory and OpenMPparentComments−imtringued2yExplicitly 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.
Comments
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.