The world’s fastest supercomputers now run more than a quintillion calculations a second

Somewhere inside a government laboratory in Tennessee sits a machine that can perform more arithmetic in one second than the world’s population could complete in thousands of lifetimes. Frontier, built for the Department of Energy’s Oak Ridge National Laboratory, became the first computer ever confirmed to sustain more than a quintillion calculations per second, crossing a threshold engineers had chased for more than a decade. Its arrival opened what the computing industry now calls the exascale era.

Oak Ridge’s Race to Exascale

Exascale computing refers to systems capable of performing at least one quintillion, or 10 to the 18th power, floating-point operations per second, a benchmark commonly abbreviated as an exaFLOP. Reaching that milestone had been a stated goal of the Department of Energy’s Exascale Computing Project for more than a decade before Oak Ridge, working with Hewlett Packard Enterprise’s Cray division and chipmaker AMD, deployed Frontier starting in 2021 and brought it to full capability in 2022.

The achievement was not simply a matter of adding more processors. Existing projections before Frontier’s construction suggested reaching a single exaFLOP might require hundreds of thousands of graphics processors and total power consumption between 150 and 500 megawatts, a level of energy draw that would have made the machine impractical to build or operate. Engineering the system to hit exascale performance within a realistic power budget became as much of a challenge as the raw computing target itself.

Inside Frontier’s Processing Power

Frontier’s measured performance reached 1.102 exaFLOPS, or 1.102 quintillion floating-point operations per second, when it first topped the industry’s TOP500 rankings in June 2022, with later hardware refinements pushing its sustained performance to 1.353 exaFLOPS and its theoretical peak to roughly 2.055 exaFLOPS. That output comes from 9,472 AMD Epyc “Trento” processors totaling more than 606,000 CPU cores, paired with 37,888 AMD Instinct MI250X graphics accelerators contributing more than 8.3 million additional cores.

All of that hardware is packed into 74 refrigerator-sized cabinets connected by high-speed Slingshot networking switches capable of moving 12.8 terabits of data per second, arranged so that no two computing nodes are more than three network hops apart. Roughly 145 kilometers of cabling, some optical and some copper, link the Frontier system together, custom-cut to keep every connection as short as engineering constraints allowed. Each of the machine’s compute nodes pairs one central processor with four graphics accelerators and several terabytes of local flash storage, feeding an internal storage system built to sustain tens of terabytes of read and write speed alongside a much larger site-wide file system used to hold research data between runs.

Cooling a 24-Megawatt Machine

Running hundreds of thousands of processor cores continuously generates an enormous amount of waste heat, and Frontier addresses that with liquid cooling rather than the fans used in ordinary computers. Four 350-horsepower pumps circulate roughly 6,000 gallons of water through the system every minute, a design that lets Frontier pack roughly five times the computing density into its cabinets compared with an air-cooled system of similar scale.

Even with that efficient cooling approach, Frontier draws around 21 to 24.6 megawatts of continuous power, comparable to the electricity demand of roughly 15,000 average American homes running at once. The project’s roughly $600 million construction cost, funded jointly by Oak Ridge, HPE Cray and AMD, reflects both the scale of the processing hardware and the specialized cooling infrastructure needed to keep it running reliably.

Topping — and Losing — the World Rankings

When Frontier’s 1.1 exaFLOP result was confirmed in the June 2022 edition of the TOP500 list, the industry’s twice-yearly ranking of the world’s fastest computers, it displaced Japan’s Fugaku system to become the fastest supercomputer on the planet. Frontier also briefly topped the Green500 list ranking computing efficiency, achieving more than 62 gigaflops of performance per watt of power consumed, before a smaller system built by the Flatiron Institute claimed that efficiency crown later the same year.

Frontier’s run atop the overall speed rankings lasted until November 2024, when the Department of Energy’s newer El Capitan system, also built with AMD hardware, surpassed it with a measured 1.7 exaFLOPS. That succession illustrates how quickly the exascale frontier keeps moving once the first machine proves the threshold is achievable; Frontier remains among the small handful of computers in the world capable of exascale-class performance even after losing the top spot.

What Exascale Computing Means Going Forward

Machines operating at this scale are put to work on problems that smaller systems simply cannot handle in a reasonable timeframe, including detailed climate modeling, nuclear stockpile simulations that substitute for physical testing, and the design of new materials at the atomic level. Frontier in particular has been used by researchers refining climate forecasts and developing new battery and materials chemistries, work that depends on running enormous physics simulations far faster than previous generations of hardware allowed.

With El Capitan now holding the overall speed record and other nations pursuing their own exascale projects, Frontier’s legacy rests less on being the single fastest machine forever and more on proving that the quintillion-calculation threshold was achievable within a workable power budget at all. Every exascale system built since has followed the engineering path Frontier’s designers had to figure out first.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview