AMD delivers Zen 4 supercomputer cores with 2 exaflops of computing power reken
AMD is helping to build a new supercomputer called El Capitan. The system is built around Epyc server processors with Zen 4 cores on board. El Capitan would get 2.2 exaflops of computing power, more than ten times as much as today’s most powerful supercomputer.
El Capitan is being developed for the US Lawrence Livermore National Laboratory. This government institute mainly conducts research into nuclear safety, including simulations of existing and new nuclear weapons. For such calculations, the institute currently has the Sierra supercomputer, number 2 on the Top500 list of supercomputers. AMD will supply an exascale computer for another US government agency as early as 2021; Frontier should yield about 1.5 exaflops at the Ministry of Energy next year.
El Capitan will be made with a new generation of Epyc server processors codenamed Genoa. That generation will get Zen 4 cores on board, which will be produced at 5nm and, according to AMD, will deliver significantly more IPC than the upcoming Milan processors. It is not yet known what those improvements will be, but El Capitan would become much faster than the current supercomputers. With a computing power of at least 2 exaflops, El Capitan would become ten times faster than the Summit supercomputer, the current number 1 in the Top500 supercomputers list.
AMD has not yet announced how many processor cores will be used for that computing power. It is known that the Epyc Genoa processors are connected to, among other things, GPUs via a new, third generation Infinity Fabric. The ratio of processors to GPUs would be 1:4 and the GPUs in El Capitan will consist of a new generation of Radeon Instinct cards. This Instinct generation will receive a new generation of HBM memory and will be optimized for deep learning and other HPC applications. In addition, Genoa, and the Instinct cards via the Infinity Fabric, will have a new generation of working memory, probably ddr5.
While AMD has not yet announced how many nodes El Capitan will consist of, it has indicated that it will consume significantly less than 40MW of power. Current supercomputers require 10 to 20MW of energy, which would imply that El Capitan burns significantly more. The energy requirement would also be closer to 30MW, which is the capacity of Cray’s Shasta supercomputing platform. Shasta uses water cooling to dissipate the heat and features Slingshot interconnects to connect the nodes together.
El Capitan involves an investment of $600 million and is expected to be completed in early 2023. The Zen 4 Genoa-generation Epyc processors are expected as early as 2021. This year AMD comes with the Milan-generation Epyc with Zen 3 cores on board.