HP, AMD’s El Capitan Becomes the World’s Fastest Supercomputer

HP and AMD have claimed the top spot on the Top500 project’s list of the world’s best-performing supercomputers, with El Capitan at the Lawrence Livermore National Laboratory in California achieving a peak High-Performance Linpack score of 1.742 Exaflops. An Exaflop is defined as one quintillion (10¹⁸) double-precision floating point operations per second). The El Capitan supercomputer was built by HP using AMD’s 4th Gen Epyc datacentre CPUs and Instinct MI300A APUs, also called accelerators. 

First announced in 2020 with a 2023 delivery target, the new supercomputer came online in mid-2024 after nearly a year of construction. It replaces IBM’s 125-Petaflop Sierra supercomputer from 2018 providing over 20 times the compute power, and will be deployed for the National Nuclear Security Administration, which manages the development and safety of the USA’s nuclear energy assets. It will be used to safely simulate nuclear weapons tests as well as materials discovery, seismic modelling, and training AI-based workflows for US national security. It was projected to be more powerful than the top 200 supercomputers in the world (as of 2020) combined. A target of 2 Exaflops was previously announced, and it is not yet clear whether work is still on to achieve this.

Although the exact final specifications have not been announced, El Capitan is known to be based on AMD’s 4th Gen Epyc CPUs, codenamed “Genoa”, with 24 cores each based on the Zen4 microarchitecture. The Instinct MI300A APU combines 24 Zen 4 cores with four CDNA 3 GPU cores, and 128GB of unified HBM3 memory. It will have over 11,000 compute “nodes” made up of clusters of these chips, with over 5.43 Petabytes of memory in total. El Capitan also uses HP Enterprise’s Slingshot network switches.  

The announcement puts AMD at the top of the Top500 list for six consecutive iterations. There are now three known and tested Exaflop-scale supercomputers in the world, all of them in the USA. While China is widely reported to have developed at least two of its own, no test results have been submitted to Top500, and neither specifications nor performance figures have been disclosed. 

The second place on the list goes to AMD’s Frontier system (1.35 Exaflops) at the Oak Ridge National Laboratory while HPE also built the Aurora supercomputer (1.01 Exaflops) for the Argonne National Laboratory using Intel hardware. Nvidia hardware tops out at 4th place, in Microsoft's Eagle supercomputer which powers its Azure infrastructure. 

India's highest ranking is now 136, with AIRAWAT-PSAI, operated by the Centre for Development of Advanced Computing in Pune for use by eligible academic and research facilities as well as startups under the National Programme for AI. 

El Capitan also reached the 18th spot on the most recent Green500 list which measures power efficiency, with its power consumption rated at 29,581kW – described by the lab as “enough power to run a mid-sized city”. Two smaller sibling supercomputers, called Tuolumne and RZAdams, ranked 10 and 49 on the new Top500 list, share the same basic architecture. Tuolumni will be used for unclassified projects in energy security, medical research, climate modelling, and other areas of public interest, while RZAdams will also be used for weapons research, according to the Lawrence Livermore National Laboratory. 

Interestingly, three more AMD supercomputers on the Top500 list – called Tenaya, Tioga, and RZVernal, and ranked  at 199, 222 and 299 respectively – are actually development test platforms for El Capitan.