Lawrence Livermore National Laboratory announced Wednesday that it has received the first components of its upcoming supercomputer, El Capitan, and has begun installation. The system is expected to go live in mid-2024 and is expected to perform in excess of 2 ExaFLOPS.
LLML’s El Capitan is based on Cray’s Shasta supercomputer architecture and, like its two other exascale systems in the United States (Frontier and Aurora), will be built by HPE. Unlike the first two exascale machines, which used traditional discrete CPU and discrete GPU configurations, the El Capitan supercomputer uses AMD server-grade CPUs that combine both processor types into a single, highly connected package. It will be the first APU-based supercomputer.
AMD’s Instinct MI300A APU incorporates both CPU and GPU chiplets, offering 24 general-purpose Zen 4 cores, a compute GPU with CDNA 3 architecture, and 128 GB of integrated on-package HBM3 memory. increase. AMD has been evaluating the Instinct MI300A APU internally for several months. AMD and HPE now appear to be ready to begin installing the first hardware that will make up El Capitan.
Engineers at the Lawrence Livermore National Laboratory have already installed quite a few servers in racks, according to a photo released by Lawrence Livermore National Laboratory. LLNL’s announcement remains unclear if these are “finished” servers with production quality silicon or pre-production servers that will be loaded with production silicon at a later date. Notably, Aurora’s parts were originally assembled with a prototype CPU, but in the last few months he’s only been replaced with a Xeon CPU Max chip. Given the amount of validation work required to launch a world-class supercomputer, AMD and HPE may be following a similar strategy here.
“We have started receiving and installing components for El Capitan, the first #exascale #supercomputer.” Tweet Read by LLNL. “It will still be some time before we deploy for national security purposes in 2024, but we are excited to see our long-standing commitment become a reality.”
LLNL expects El Capitan to become the world’s fastest supercomputer when it goes live in 2024. Its full specs haven’t been published yet, but it’s not clear how fast it will be on paper compared to the 2 EFLOPS Aurora, let alone its actual performance. One of AMD’s design goals for his MI300A APU is to take advantage of the further performance efficiency gains that come from having the CPU and GPU blocks very close together. So it will be interesting to see what software development teams programming specifically for El Capitan can achieve. Further optimize the software.
LLNL’s El Capitan is expected to cost $600 million to build. The system will be used to simulate nuclear weapons and will be critical to US national security. It replaces Sierra, a supercomputer based on IBM Power 9 and NVIDIA Volta accelerators, and promises 16x higher performance.