Supercomputer

Supercomputers are high-performance computing systems designed to execute exceptionally large numbers of calculations at speeds far beyond those of general-purpose computers. Their performance is measured primarily in floating-point operations per second (FLOPS), a metric that reflects the mathematical intensity of the tasks they handle. Since 2022, the world has entered the era of exascale computing, with supercomputers capable of exceeding 10¹⁸ FLOPS. By comparison, a typical desktop computer performs in the range of hundreds of gigaflops to tens of teraflops, underscoring the enormous gap between consumer and supercomputing performance. Modern supercomputers overwhelmingly operate on Linux-based systems and are developed globally, with major initiatives in the United States, the European Union, China, Japan, Taiwan, and other technologically advanced regions.
Supercomputers are indispensable in computational science, supporting high-intensity applications such as quantum mechanics, climate modelling, weather prediction, computational chemistry, nuclear simulation, aerospace design, large-scale astrophysical modelling, and cryptanalysis. Their capacity to perform trillions of simultaneous operations allows researchers to simulate complex physical systems that are impossible to reproduce experimentally.

Early development and the rise of high-performance computing

The roots of supercomputing reach back to the 1960s. One of the earliest machines regarded as a supercomputer was the UNIVAC LARC, designed for the United States Navy Research and Development Center. It employed high-speed drum memory rather than the more modern disk technologies then being developed. Around the same time, IBM produced the IBM 7030 Stretch for the Los Alamos National Laboratory, which had requested a machine one hundred times faster than existing computers. Although the Stretch did not fully meet this goal, it introduced important advances, including transistor-based construction, magnetic core memory, pipelined instruction handling, and early random-access disk storage. It laid the foundation for the IBM 7950 Harvest, built for cryptanalytic applications.
Another influential project of the era was the Atlas computer, designed by Tom Kilburn at the University of Manchester. Atlas integrated magnetic core memory with drum storage and implemented one of the earliest examples of page-based memory swapping through the Atlas Supervisor. It also introduced the innovative concept of timesharing, allowing multiple programs to run concurrently. With speeds approaching a microsecond per instruction, Atlas was among the most advanced computing systems of its time.
A major turning point occurred with the CDC 6600, completed in 1964 by Seymour Cray. It replaced germanium transistors with faster silicon devices and used refrigeration to maintain stable operation. The CDC 6600 exceeded the performance of all contemporary machines by a factor of ten and effectively defined the category of the supercomputer. Approximately one hundred units were sold, each at a premium price, cementing its place in computing history.
Cray’s departure from CDC in 1972 led to the founding of Cray Research, which produced some of the most successful supercomputers of the twentieth century. The Cray-1, released in 1976, became iconic for its unique design and vector-processing capabilities, which dominated supercomputing architecture for decades. Its successor, the Cray-2 (1985), introduced liquid cooling using Fluorinert and became the first system to achieve speeds above one gigaflop.

Vector processing and the transition to parallelism

Throughout the 1970s and 1980s, vector processors were central to supercomputer design. These machines handled operations on large arrays of data with extraordinary efficiency, making them ideal for scientific workloads. However, the concept of parallelism—using many processors simultaneously—emerged as another important direction.
A pioneering example was the ILLIAC IV, conceived in the late 1960s as a massively parallel system with 256 processors. Difficulties in development resulted in only 64 processors being built, and the machine ultimately achieved lower performance than expected. However, its partial success demonstrated the potential of parallel computing. Seymour Cray famously expressed scepticism towards the approach, comparing the choice between two powerful processors and thousands of weaker ones to choosing between two strong oxen and many chickens for ploughing a field.
By the early 1980s, however, massively parallel architectures gained traction. MIT’s Connection Machine series exemplified this trend: the CM-1 included up to 65,536 custom microprocessors connected in a sophisticated network. Later models such as the CM-5 enabled billions of arithmetic operations per second. Parallel computing efforts also flourished in Japan, where Osaka University’s systems used hundreds of microprocessors to generate realistic three-dimensional graphics.
Fujitsu’s VPP500 (1992) introduced high-speed gallium arsenide processors, while the company’s Numerical Wind Tunnel system dominated supercomputing in 1994 with vector processors capable of 1.7 gigaflops per processor. Other major contributions included Hitachi’s SR2201 (1996), which used 2,048 processors linked through a fast three-dimensional crossbar network, and Japan’s CPPACS project, which further advanced massively parallel design for scientific simulation.

Modern architectures and exascale computing

From the late 1990s onward, supercomputers increasingly adopted massively parallel designs using tens of thousands of commodity processors rather than custom hardware. These architectures supported vast compute clusters linked by high-speed interconnects. A representative example is the Intrepid supercomputer, installed at Argonne National Laboratory and composed of 164,000 processor cores arranged in 40 racks using a torus interconnect. Such systems rely on sophisticated cooling and power distribution but often use standard data-centre cooling rather than bespoke refrigeration.
The twenty-first century has seen rapid advances in supercomputing capability. Many of the world’s fastest machines now exceed hundreds of petaflops, and by 2022 the first exascale-class supercomputers became operational. The United States retains a leading position in the TOP500 rankings through systems developed by national laboratories, such as Lawrence Livermore National Laboratory’s El Capitan, one of the leading exascale systems. Japan, China, Italy, Finland, and Switzerland also host high-ranking machines, with China believed to operate exascale systems that remain unlisted on public rankings.
The global progression of supercomputers reached a milestone in 2018 when the total combined performance of systems on the TOP500 list surpassed one exaflop.