In November 2024, the AMD-powered El Capitan claimed the title of the world’s fastest supercomputer, achieving a peak performance of 2.7 exaflops and sustaining 1.7 exaflops. Designed by HPE for the National Nuclear Security Administration (NNSA), this system resides at the Lawrence Livermore National Laboratory (LLNL) and is tasked with simulating nuclear weapons tests. Its advanced computing power is driven by AMD Instinct MI300A APUs, dethroning Frontier to become the most powerful supercomputer globally.
Table of Contents
Behind the Scenes at LLNL
Patrick Kennedy of ServeTheHome attended the unveiling of El Capitan at LLNL in California. The event featured AMD and HPE’s CEOs, alongside a rare opportunity to glimpse the system before it commenced its classified operations.
El Capitan’s infrastructure features 128 compute blades per rack, each liquid-cooled for optimal performance and noise reduction. Kennedy remarked that the surrounding storage systems produced more noticeable sound compared to the quiet operation of the blades.
The system integrates HPE’s Slingshot interconnect, combining DAC and optical cables for seamless communication. Although only half of the interconnect space is populated with switch trays, the remaining space houses the “Rabbit,” a liquid-cooled unit equipped with 18 NVMe SSDs to support specific computing requirements.
Hardware Insights
Kennedy observed an AMD EPYC 7003 Milan processor within the system, which aligns with the generation of AMD’s MI300A APUs. The Rabbit’s CPU includes DIMMs and DDR4 memory, all cooled with liquid systems. Remarkably, the design eliminates the need for fans, making the entire setup liquid-cooled.
While El Capitan is smaller than xAI’s Colossus cluster—equipped with 100,000 Nvidia H100 GPUs as of September—Kennedy noted its efficiency. Despite its relatively compact footprint, El Capitan demonstrates significant performance at a fraction of the budget required for a massive GPU-based system.
