Frontier becomes the first ExaFLOP supercomputer: built with AMD EPYC & Instinct MI250X

Over the years, supercomputers have been pushing various computational limits, including MegaFLOP, GigaFLOP, TeraFLOP, PetaFLOP, and now ExaFLOP. Recently, TOP500 officially announced the 59th list, the Frontier system at Oak Ridge National Laboratory (ORNL) in the US has become the first supercomputer to reach the ExaFLOP level in a true sense. The High-Performance Linpack (HPL) benchmark showed a computing performance of 1.102 Exaflop/s and a peak performance of 1.685 Exaflop/s.

Frontier supercomputer is based on the HPE Cray EX235a architecture, using AMD’s third-generation EPYC processor with a clock of 2GHz, and the entire system has a total of 8,730,112 cores. According to past information, each HPE Cray EX node includes a 64-core EPYC Trento 7A53 processor from AMD, 512GB of DDR4 memory, and four Instinct MI250X compute cards, connected via four HPE Slingshot 200GBps Ethernet NICs (25GB/s), provide 800Gbps of node bandwidth.

The EPYC Trento processor is a derivative of the Zen 3 architecture code-named Milan. It is rumored that its I/O chip uses Infinity Fabric 3.0 to achieve a memory interface consistent with the GPU. Each EPYC Trento processor is divided into four NUMA regions, and each NUMA region is connected to an Instinct MI250X compute card (two GCDs each). The CPU and GPU are connected through the Infinity Fabric with an interface bandwidth of 36+36GB/s, and the total bandwidth of 288GB/s between the CPU and the GPU is distributed among the eight GCDs in the node.
In the past several lists, the top spot has been held by the Fugaku supercomputer built by Fujitsu and installed at the RIKEN Computational Science Center (R-CCS) in Kobe, Japan, using the A64FX processor. This time, Fugaku slipped to second place, its HPL performance is 442 Petaflop/s, and its peak performance can theoretically reach 1 Exaflop/s. Finland’s CSC’s EuroHPC Center ranked third in terms of LUMI, a new supercomputer on the list, with an HPL performance of 151.9 Petaflop/s.