High-Performance & AI Computing

Built at the University of Pennsylvania, the machine could be reprogrammed for different calculations rather than built for a single task, and its wartime work computing artillery firing tables made it the direct ancestor of every later high-performance computing system.

Designed by Seymour Cray for Control Data Corporation, the machine is widely regarded as the first true supercomputer, using parallel processing units to outperform every other computer of its era by a wide margin.

Seymour Cray's distinctive cylindrical machine used vector processing to outperform conventional computers by a wide margin, established Cray as the industry's defining brand for a generation and set expectations for what a dedicated supercomputer should be able to do.

Compiled twice a year using a standard benchmark, the list gave governments, laboratories and companies a common, comparable measure of computing power and turned the race to build the fastest machine into a closely watched public contest.

Built under the Accelerated Strategic Computing Initiative to simulate the reliability of the American nuclear arsenal without live testing after the Comprehensive Test Ban Treaty, the machine tied the country's most sensitive weapons programme directly to the frontier of computing performance.

The climate-modelling machine outperformed the next fastest system in the world by several times, a gap large enough to prompt an anxious reassessment in the United States of its own commitment to high-performance computing.

From around 2005 the dominant architecture for high-performance computing moved away from monolithic supercomputers towards clusters and grids of commodity machines. This reorganised how nations and institutions assembled computational power, favouring aggregation of many nodes over single bespoke systems.

Installed at the United States Department of Energy's Los Alamos National Laboratory, IBM's Roadrunner combined conventional processors with chips derived from video game hardware to become the first computer to sustain one petaflop, crossing a performance threshold years ahead of most forecasts and showing how commodity components could be harnessed for national laboratory workloads. Ranked among the world's most powerful supercomputers, it was used to simulate the performance, safety and reliability of nuclear weapons and to certify their functionality, illustrating the strategic security role of national HPC capacity.

The system's rise to first place marked the first time a Chinese supercomputer led the global rankings, signalling the country's arrival as a serious competitor in a field the United States and Japan had dominated.

Built entirely with Chinese-designed processors rather than imported American chips, the machine demonstrated that China could compete at the very top of supercomputing without relying on foreign semiconductor technology, a capability that took on greater importance as export controls tightened in later years.

The US Department of Energy system exceeded one exaflop of sustained performance on the standard benchmark, a threshold researchers had targeted for over a decade, and reopened a period of intense international competition to build ever faster machines.

The kingdom added computing capacity intended to support scientific research, weather and climate modelling and a wider strategy of positioning the country as a regional hub for advanced computing and artificial intelligence infrastructure.

SK hynix and Samsung scaled up manufacturing of the memory chips that sit alongside processors in the most powerful computing systems, a component in which South Korean firms held a commanding share of global supply.

A broad rule restricting where the most advanced AI and HPC chips could be sold, issued in the final days of the previous administration, was rescinded and replaced with a narrower approach, leaving allies and chipmakers uncertain about the rules for months.

Installed at the Julich research centre under the EuroHPC joint programme, the machine gave European researchers exascale computing capacity for climate modelling, materials science and other work that had previously required using systems in the United States or Japan.

Additional PARAM-series systems were installed at universities and research institutes, and the government backed development of indigenous processors, aiming to reduce dependence on imported hardware for scientific computing and to widen access beyond a handful of national laboratories.

The project, developed with American technology firms and the Emirati company G42, aimed to build one of the largest concentrations of advanced computing capacity outside the United States, positioning the Gulf state as a hub for regional and international compute demand.

Riken and Fujitsu worked with an American chipmaker on the follow-on system, intended to keep Japan among the leading nations in supercomputing after Fugaku's early lead in world rankings faded as other countries reached exascale performance.

Chinese laboratories were widely believed to operate machines that would rank among the fastest in the world, but stopped submitting results to the international TOP500 list years earlier, while the government pressed ahead with domestic processors to reduce dependence on foreign chips.

Energy company Eni brought HPC6 into operation in Italy in November 2024, reaching 606 petaFLOPS. It marked the peak of privately owned industrial HPC capacity, applied to energy research rather than state science programmes.

Member states agreed a framework touching on artificial intelligence governance and the unequal distribution of computing capacity, an acknowledgement that most of the world's high-performance computing power remained concentrated in a small number of countries.

The US Department of Energy system, built around AMD processors and graphics chips, topped the global TOP500 ranking with peak performance above 1.7 exaflops and was dedicated primarily to simulating the reliability of the American nuclear weapons stockpile.