El Capitan supercomputer began to install AMD Instinct MI300 APU

The United States’ Lawrence Livermore National Laboratory (LLNL) has announced that the ExaFLOP-scale supercomputer, El Capitan, is currently undergoing component installation, with plans for a comprehensive launch in 2024. El Capitan will feature multiple nodes, each equipped with several Instinct MI300 APUs, to be installed on the new SH5 sockets (LGA 6096). To ensure seamless hardware and software operation, El Capitan likely underwent the initial phase of installation approximately a year ago.

Similar to Frontier and Aurora, El Capitan is based on Hewlett Packard’s Shasta supercomputer architecture. Consequently, manufacturing responsibility lies with Hewlett Packard Enterprise (HPE), employing the Slingshot-11 interconnect to link each HPE Cray XE rack. Upon completion around mid-2024, El Capitan’s FP64 peak computing performance is expected to reach 2 ExaFLOPs, rendering it the world’s fastest supercomputer. This represents a tenfold increase from its predecessor, Sierra, which began operations in 2018 with IBM’s Power9 CPU and Nvidia’s Volta architecture GPU.

El Capitan is the first to utilize an APU as the core of a supercomputing system, paving the way for AMD’s future Exascale APU model development. The Instinct MI300 APU will be fabricated using a 5nm process, featuring 24 CPU modules with the Zen 4 architecture, a GPU module with the CDNA 3 architecture, complemented by 128GB of HBM3, and equipped with next-generation Infinity Cache. It will also utilize the fourth-generation Infinity architecture, supporting the CXL 3.0 ecosystem.

AMD has already conducted several months of internal testing on the Instinct MI300, appearing ready to deliver to key customers, with an expected market release by the end of 2023.