Multi-GPU systems are widely exploited today by large-scale datacenter / ML systems. The Multi-Chip Module GPU (MCM GPU) may become the future of GPUs since the slowdown of transistors scaling on the monolithic-chip GPU. AMD MI250X GPU built on top of CDNA2 architecture is the first available commercial MCM GPU on the market and deployed in the first public exascale supercomputer –– Frontier.
Papers
- [ISCA ‘17] MCM-GPU: Multi-Chip-Module GPUs for Continued Performance Scalability