El Capitan: Inside America’s Most Powerful Supercomputer

Science & A.I. · Exascale computing

For nineteen months El Capitan was the fastest computer humanity had ever built. On 23 June 2026 it lost that title — and the people who run it were entirely unbothered, because the number-one ranking was never the point.

The machine lives in a windowless hall at Lawrence Livermore National Laboratory in California, drawing roughly the electricity of a small town. It was built not to win benchmarks but to answer one consequential question without setting off a nuclear explosion: are the weapons in the United States stockpile still safe, and do they still work?

El Capitan exists to simulate that answer. This is the story of what an exascale supercomputer actually is, why a weapons laboratory needed the most powerful one on Earth, and what it means that a Chinese system has just overtaken it.

What “exascale” actually means

An exaflop is a quintillion floating-point operations per second — a one followed by eighteen zeros. The scale defeats intuition, so try this. If every person alive completed one arithmetic sum every second, without pause, it would take more than four years of unbroken human effort to match what El Capitan does in one second.

The benchmark that certifies this is called High Performance Linpack, or HPL. It measures how fast a system solves a vast dense set of linear equations, and its result — a figure called Rmax — is what the twice-yearly TOP500 list ranks. When El Capitan debuted in November 2024 it posted 1.742 exaflops on HPL.

That made it the third machine ever to cross the exascale line, and at that moment the fastest of them all. Rmax is the measured speed; a second figure, Rpeak, is the theoretical ceiling the hardware could reach if nothing ever waited on anything else.

El Capitan’s Rpeak is 2.79 exaflops, so its real-world HPL result runs at roughly 62 per cent of the paper maximum. That ratio is respectable for a machine of this size, where the gap between peak and delivered performance is exactly where the hardest engineering hides.

The road to exascale: how we got here

El Capitan did not appear from nowhere. It is the product of a decade-long United States Department of Energy programme to reach exascale computing, and it stands on the shoulders of the machines that came before it.

The exascale era began in earnest with Frontier, which came online at Oak Ridge National Laboratory in 2022 as the first system to break the exaflop barrier on HPL. Frontier, another HPE Cray machine, proved that sustained exascale performance was achievable and set the template that El Capitan would follow and surpass.

Livermore had its own lineage. Its previous flagship, Sierra, delivered around a tenth of El Capitan’s speed and served the same stockpile mission for years. Each generation lets the weapons laboratories simulate more physics, in higher resolution, in less time — a steady climb that El Capitan advanced by a large step.

Crucially, El Capitan is described as the first exascale system dedicated to national security. Where Frontier and Argonne’s Aurora are open-science machines, El Capitan was designed from the outset around the classified needs of the National Nuclear Security Administration, which shaped everything from its architecture to the hall that houses it.

Inside the machine: eleven million cores and a clever chip

El Capitan is an HPE Cray Supercomputing EX system built around more than 11,000 compute nodes and 11,039,616 total cores. Its performance comes from a component that reflects the dominant idea in modern high-performance computing: stop treating the processor and the graphics accelerator as separate chips that must ship data back and forth.

The system is powered by AMD Instinct MI300A units, described as accelerated processing units, or APUs. Each MI300A fuses AMD fourth-generation EPYC CPU cores, GPU compute cores and high-bandwidth memory into a single shared package.

Because the CPU and GPU address the same pool of memory, the costly copying of data between separate CPU and GPU memories — a chronic tax on older designs — largely disappears. For the memory-hungry physics codes the weapons laboratories run, that unified design is not a minor convenience; it is much of the reason the machine performs as it does.

Those nodes are stitched together by the HPE Slingshot 11 interconnect, the nervous system that lets eleven million cores behave as one machine rather than eleven million isolated ones. In a calculation spread across the whole system, the speed at which nodes exchange partial results often matters as much as raw core count.

The whole assembly carries 5.4375 petabytes of memory and is kept alive by direct liquid cooling, water piped straight to the hot components rather than relying on chilled air. At full load the machine draws about 30 megawatts.

Efficiency, not just raw speed, is a deliberate goal. At its debut El Capitan delivered 58.89 gigaflops for every watt consumed, a figure that had improved to roughly 60.9 gigaflops per watt by late 2025 as its software and operation matured — placing it among the more efficient large systems in the world.

Why a weapons lab needed the world’s fastest computer

The United States has not conducted an explosive nuclear test since 1992. That moratorium created a scientific problem that has quietly shaped American supercomputing ever since: how do you know an ageing arsenal still works, and is still safe to handle, if you never detonate one to check?

The answer is the National Nuclear Security Administration’s Stockpile Stewardship Program, which replaces physical testing with physics-based simulation. El Capitan is its most powerful instrument, a shared resource for the three NNSA weapons laboratories — Lawrence Livermore, Sandia and Los Alamos.

The physics involved is genuinely hard. Certifying a warhead means modelling how materials behave under extreme pressure and temperature, how components age over decades, and how a vast number of coupled physical processes unfold in fractions of a second. Each added dimension of realism multiplies the computing required.

Jill Hruby, the NNSA Administrator and Department of Energy Under Secretary for Nuclear Security, framed El Capitan’s arrival as the continued advance in capability needed to sustain the stockpile without resuming explosive testing.

Rob Neely, who directs weapon simulation and computing at Livermore, has described the goal in similar terms: simulating entire weapons systems while incorporating the many real-world factors a simpler model is forced to leave out. Each of those factors adds computational cost, and the machine exists to absorb it.

Simulation does not stand alone. It is calibrated against the archive of data from the era of underground testing, against non-nuclear subcritical experiments, and against results from facilities such as the National Ignition Facility. The computer ties those threads together into a predictive model, which is why more computing power translates directly into more confidence.

This is why the leaderboard ranking is close to irrelevant to the laboratory. A weapon simulation that once had to run in coarse approximation, or in two dimensions to save time, can now run in full three-dimensional detail with more of the underlying physics included. The value is measured in the fidelity of the answer, not a position on a public list.

Beyond weapons: what else exascale unlocks

A machine built for weapons physics turns out to be extraordinarily useful for any problem that is too large, too dangerous or too slow to study by experiment alone. Livermore is also home to the National Ignition Facility, where researchers achieved fusion ignition in 2022, and the same laboratory culture of simulation-led science runs through both projects.

The ability to resolve fine detail across an enormous simulated volume applies far beyond the stockpile mission. Climate and weather models can be run at finer resolution, capturing smaller features of the atmosphere and ocean. Drug and materials discovery can screen enormous numbers of candidate molecules before anyone synthesises one in a laboratory.

Fusion-energy research, cardiac modelling and the simulation of new battery chemistries all draw on the same capacity to compute the consequences of physics at scale. Problems that were once reduced to rough approximations can be attacked in something closer to their full complexity.

Increasingly these machines also run artificial-intelligence workloads. The MI300A accelerators that give El Capitan its speed are the same class of hardware used to train large neural networks, so an exascale system can move between traditional physics simulation and modern machine learning — sometimes weaving the two together, using a trained model to approximate an expensive calculation and free the simulation to reach further.

What the benchmark measures — and what it misses

It is worth being precise about what a TOP500 ranking does and does not tell you. HPL solves one particular kind of mathematical problem extremely efficiently, and it rewards machines with enormous dense-matrix throughput. It is a fair, repeatable yardstick — but it is a single number standing in for a machine that does far more.

Many real scientific codes look nothing like HPL. They move irregular data, communicate constantly between nodes, and stress memory and network in ways the benchmark does not. This is why the supercomputing community also tracks alternative measures, and why a laboratory judges a machine by whether its own applications run well, not by a headline figure.

The rise of artificial intelligence has added another dimension. AI training leans on lower-precision arithmetic, where a machine can run far faster than its double-precision HPL score suggests. A system’s single TOP500 number, in other words, is the beginning of its story rather than the whole of it.

Dethroned: the June 2026 TOP500 upset

El Capitan held first place through three editions of the TOP500 list. In November 2025 it was still comfortably ahead, its HPL result having climbed to 1.809 exaflops as the system settled into full operation. Then, on 23 June 2026, at the ISC High Performance conference in Hamburg, the order changed.

A previously unlisted Chinese system called LineShine, installed at the National Supercomputing Centre in Shenzhen, debuted at number one with an HPL score of 2.198 exaflops against a theoretical peak of 2.736 exaflops. It became the first machine on the list to exceed two exaflops of sustained double-precision performance using CPUs alone, without GPU accelerators.

That CPU-only approach is a notable departure from the accelerator-heavy designs that dominate the rest of the top tier, and its appearance is significant for another reason: Chinese centres had for years declined to submit systems to the TOP500 at all. A debut straight at number one signals both renewed transparency and considerable capability.

El Capitan moved to second at its steady 1.809 exaflops. Behind it sat the other American exascale machines, Frontier at Oak Ridge with 1.353 exaflops and Aurora at Argonne with 1.012 exaflops, followed by Europe’s first exascale system, the JUPITER Booster at the Jülich Supercomputing Centre in Germany, at an even 1.000 exaflops.

The reshuffle is a reminder that a supercomputer ranking is a snapshot of a moving race, not a settled fact. El Capitan lost none of its capability on 23 June; a faster machine simply appeared. For the geopolitics of computing, though, it matters a great deal that the top of the list is now contested across three continents.

The US–China computing race behind the numbers

LineShine’s debut is about more than one benchmark. For several years, leading Chinese centres had stopped submitting their fastest machines to the TOP500, even as researchers believed exascale systems were operating inside the country. A public return straight to number one is a deliberate statement.

The detail that it reached two exaflops using CPUs alone, without the GPU accelerators that power El Capitan, Frontier and Aurora, is strategically pointed. The United States has restricted exports of the most advanced accelerator chips to China, and a competitive CPU-only design is one answer to those controls.

For the United States, the response is that leadership was never resting on a single machine. The country now operates three exascale systems — El Capitan, Frontier and Aurora — spanning weapons science and open research, a depth of installed capability no other nation yet matches. Europe’s JUPITER has joined the exascale club, and the field is widening rather than narrowing.

The lesson of the June 2026 list is that exascale computing is no longer a two-horse race, and no single ranking will stay settled for long. National capability is better measured by how many frontier machines a country can build, run and use well — not by who happens to hold the top line in any given month.

What it takes to build and run an exascale machine

A system like El Capitan is not bought off a shelf. It is the product of years of co-design, in which the laboratory, the system builder Hewlett Packard Enterprise and the chip designer AMD shape the hardware and the scientific software together, so that the machine fits the problems it must solve rather than the other way round.

The hardest part is often invisible. Building a machine with eleven million cores is one challenge; writing software that keeps all of them usefully busy is a far greater one. A calculation must be broken into millions of pieces that run in parallel, exchange results constantly, and still arrive at a single coherent answer.

Decades-old physics codes had to be substantially rewritten to run efficiently on the MI300A architecture, a preparation effort that ran for years before the hardware was switched on. Much of the return on an exascale investment comes not from the metal but from that painstaking work of readying applications to exploit it.

Then there is the sheer physical reality of the thing. A machine drawing around 30 megawatts needs the power feeds, the water-cooling plant and the reinforced floor of a small industrial facility. Running it is a continuous engineering operation, and the energy appetite of frontier computing is now part of a wider debate about the sustainability of large-scale data centres.

The real measure of a machine like this

It is tempting to treat these systems as trophies, and press coverage tends to reward that framing. But the engineers at Livermore will tell you the ranking is the least interesting thing about El Capitan. What matters is that a class of scientific question that was previously out of reach can now be answered in full.

The deeper significance of exascale is that it changes what counts as a knowable problem. A machine that can simulate an ageing warhead in three dimensions can also model a beating heart, a warming ocean or a fusion plasma with the same unprecedented fidelity.

El Capitan may no longer be the fastest computer on Earth. But the era it helped open — where simulation stands in for experiments we cannot, or must not, perform — is only beginning. The leaderboard will keep changing hands; the deeper shift, in what science can now compute, is the part that lasts.

Frequently asked questions

Sources

Primary institutional sources:

• Lawrence Livermore National Laboratory — El Capitan verified as world’s fastest supercomputer (2024)
• LLNL Advanced Simulation and Computing — El Capitan: NNSA’s first exascale machine
• Hewlett Packard Enterprise — HPE delivers El Capitan for LLNL (Nov 2024)

Rankings and science journalism:

• TOP500 — El Capitan system entry (DOE/NNSA/LLNL)
• TOP500 — LineShine debuts at No. 1 (June 2026 list)
• TOP500 — El Capitan retains No. 1 as JUPITER becomes Europe’s first exascale system (Nov 2025)
• Oak Ridge Leadership Computing Facility — Frontier and the dawn of the exascale era