AMD, the Instinct MI200 accelerator: CDNA 2 and MCM

After announcing the Instinct MI100 accelerator last November, the first based on CDNA (Compute DNA) architecture, AMD could introduce a new model called Instinct MI200 this year. The company would seem to want to impress an acceleration of efforts to create a valid alternative to Nvidia’s proposals in the HPC sector, presenting a new solution one year after the previous model, while traditionally we have seen a two-year cadence.

The new solution should be based on CDNA architecture 2 and, above all, be based on a project MCM (multi-chip module), ie no longer on a “monolithic” GPU but several chips connected together on a common substrate (how and in what way is not yet known). As for the specifications, at the moment there is no concrete information, but someone advances the hypothesis that the transition from MI100 to MI200 indicates a doubling of the capacity compared to the accelerator currently available.

The new AMD Instinct MI200 should be presented in an OAM (open-source accelerator infrastructure) format, different from the card format you see in the stock image above, and debut in the supercomputer Frontier. The latter is a system that AMD is building in collaboration with Cray for the US Department of Energy (Oak Ridge National Laboratory) and which should guarantee a power greater than 1.5 exaflops.

Inside the new supercomputer AMD could also adopt some unpublished EPYC CPUs called “Trento”. The specifics of Trento are not clear at the moment, but it should not be a Zen 4 project like Genoa, but a variant of Milan (Zen 3) optimized for the world of supercomputers. Even in this case, however, it is simple speculation, there is no concrete information on Trento at the moment.