Memory requirements stemming from advanced datasets for AI and HPC applications often swamp the memory built into a GPU. Expanding that memory has typically meant installing expensive high bandwidth memory, which often introduces changes to the existing GPU architecture or software.