Cray is building an exascale cluster for the US Department of Energy in partnership with AMD. In 2021, the system should be delivered with a computing power of 1.5 exaflops. Specially designed Epyc processors and Radeon Instinct GPUs are used.
The new exascale cluster will be located at Oak Ridge National Laboratory and will be named Frontier. AMD and Cray expect it to be the world’s fastest supercomputer upon delivery. The total costs for the system are more than 600 million dollars, converted about 536 million euros.
The cluster covers approximately 678 square meters and is built using Cray’s Shasta platform. The Frontier is equipped with processors and GPUs developed by AMD. AMD speaks in a presentation about modified versions of its Epyc server processors, based on a ‘future’ Zen architecture and optimized for supercomputers.
In each node, one processor is combined with four custom Radeon Instinct GPUs, which are equipped with hbm memory. The CPUs and GPUs communicate through the Infinity Fabric interconnect. AMD does not disclose which architecture the GPUs are based on, but it is presumably Navi GPUs.
The Frontier is the second exascale supercomputer for the US Department of Energy that is being worked on. In March, Intel announced that such a system with Xeon processors and Xe GPUs, called Aurora. The Intel variant should also be released in 2021. The ministry uses the supercomputers for scientific research.
Update: The Epyc processors in the exascale computer are based on a new, yet unknown architecture, which comes after Zen 2. The article has been adapted accordingly.