It has been announced that AMD's EPYC processors will power what will be the world's fastest exascale supercomputer. Called Frontier, this new supercomputer will deliver in excess of 1.5 exaflops of compute performance and will be delivered to ORNL rough two years from now.
Although AMD is currently being a bit light on the details with regards to what specific EPYC chips will be deployed in Frontier, we do know that they will be at least Zen 2-class processors that have specific instructions added for artificial intelligence and high performance computing (HPC) at the request of (and with input from) Cray and ORNL. In other words, these will be custom chips and won't be standard off-the-shelf components available to just anyone.
The custom EPYC processors will be joined by AMD's Radeon Instinct GPU accelerators which come packing HBM2 to aid in deep learning performance. In this particular application, one EPYC processor will be coupled with four Radeon Instinct GPUs per node via AMD's Infinity Fabric interconnect. These components will be encased in each Cray Shasta Compute Blade (as seen in the image below); many of which will make up Frontier.
“Frontier represents the state-of-the art in high-performance computing," said Jeff Nichols, who serves as the associate laboratory director for Computing and Computational Sciences at ORNL. "Designing and standing up a machine of its scope requires working closely with industry, partnerships which not only enable breakthrough science but also ensure American scientific and economic competitiveness on the global stage."
AMD's second-generation EPYC "Rome" processors based on 7nm Zen 2 microarchitecture are available in up to 64-core configurations, which are capable of executing 128 threads. AMD has already showcased these processors (in early silicon form) putting the smackdown on Intel's Xeon Platinum 8180M, so it'd be interesting to see what kind of special sauce AMD has baked into these customer Zen chips to fuel Frontier.