El Capitan Supercomputer: From 2020 Announcement to Exascale

October 08, 2020

In March 2020, Hewlett Packard Enterprise (HPE) and AMD announced the hardware for El Capitan, an exascale supercomputer that the US Department of Energy's National Nuclear Security Administration (NNSA) had ordered for Lawrence Livermore National Laboratory (LLNL) in California. HPE's press release, distributed by Business Wire, said the system would reach more than 2 exaflops of peak double-precision performance. That was more than the 200 fastest supercomputers of the time combined and about ten times faster than the leader of the day, Summit at Oak Ridge National Laboratory. Delivery was planned for early 2023.

The 2020 announcement

The system itself had already been ordered. Cray, which HPE bought in 2019, won the contract of about $600 million in August 2019 with a target of roughly 1.5 exaflops. Choosing next-generation AMD EPYC "Genoa" processors with Zen 4 cores, AMD Radeon Instinct GPUs and AMD's open-source ROCm software raised the projection to 2 exaflops. HPE described this as an increase of more than 30% on the estimate made seven months earlier. That is the source of the often-quoted "30% more powerful than expected" figure: it described a projection, not a finished computer.

The main mission was clear from the start. El Capitan runs simulations for the NNSA's stockpile stewardship program, which maintains the US nuclear arsenal without explosive testing, and it serves three laboratories: Livermore, Sandia and Los Alamos. HPE's release also mentioned AI and machine learning workloads and LLNL's medical research, such as work to speed up cancer drug discovery.

How the design changed

The finished machine is different from the 2020 plan. Instead of separate processors and graphics accelerators, El Capitan uses AMD Instinct MI300A accelerated processing units (APUs). Each one puts 24 Zen 4 CPU cores, a CDNA 3 GPU and 128 GB of HBM3 memory in a single package. Because the CPU and GPU share that memory, data does not have to be copied between them, a common bottleneck in GPU computing. The APUs sit four to a node in liquid-cooled HPE Cray EX cabinets, connected by HPE's Slingshot network, for more than 11 million cores in total.

From No. 1 to No. 2

El Capitan arrived later than first planned. Installation began in 2023, and in November 2024 the system topped the TOP500 list with 1.742 exaflops on the High Performance Linpack (HPL) benchmark, taking first place from Frontier at Oak Ridge. LLNL formally dedicated El Capitan on January 9, 2025. A new measurement in November 2025 raised its score to 1.809 exaflops.

It held first place until June 2026, when LineShine, a new system at the National Supercomputing Center in Shenzhen, China, debuted at No. 1 with 2.198 exaflops. El Capitan is now No. 2. It still leads the HPL-MxP mixed-precision benchmark at 16.7 exaflops.

Why it matters beyond nuclear research

At the dedication, LLNL's Brad Wallin noted that El Capitan was bought for modeling and simulation, but said its AMD APUs and HPE technology make it "a true powerhouse for AI training and inference". The mixed-precision result shows why: AI workloads use lower-precision arithmetic, and at that precision El Capitan is about nine times faster than in its standard HPL run.

For developers and businesses, there are three practical lessons:

• Accelerators do most of the work. Eight of the ten fastest systems in June 2026 get most of their performance from GPUs or APUs, and so do the AI clusters companies rent. Performance-critical code has to be written for them. The exceptions are LineShine, which reached its score with CPUs only, and Japan's Fugaku.
• The software stack matters. El Capitan runs on AMD's ROCm rather than NVIDIA's CUDA. ROCm's HIP programming interface closely mirrors CUDA, which makes GPU code easier to port between the two platforms and reduces lock-in at a time when demand for NVIDIA's AI chips has repeatedly outstripped supply.
• You do not need your own supercomputer. Microsoft's cloud-based Eagle system is No. 7 on the June 2026 TOP500. The major cloud providers rent the same class of GPU and HPC capacity by the hour from a fast-growing network of data centers, and that is how many companies run large simulations and train AI models.