From machines you could sit on to machines powerful enough to simulate worlds
There was a time when a computer was unmistakably a computer. It occupied a room, hummed, generated enough heat to concern the engineers and was attended by specialists who understood languages largely incomprehensible to everybody else. Today, computers have almost disappeared, not because there are fewer of them, but because they are everywhere. They regulate traffic lights, route telephone calls, forecast tomorrow’s weather, process card payments, guide aircraft, manage electricity grids, help doctors interpret scans and allow billions of people to carry remarkable computing power in their pockets.
Behind that familiar world lies another, largely unseen one: supercomputers containing millions of processing cores, vast data centres working collectively, specialised processors designed for artificial intelligence and, at the experimental frontier, quantum computers whose basic units of information obey some of the strangest laws of physics. The journey from Alan Turing’s mathematical conception of a universal machine to today’s exascale computers has taken less than a century, and we may still be somewhere near the beginning.
When computers became electronic
Modern computing has many parents, and its history resists being reduced to one inventor. Alan Turing’s work in the 1930s provided one of its intellectual foundations. His theoretical “universal machine” demonstrated that a single machine could, in principle, perform any computation capable of being expressed algorithmically. During the Second World War, Turing’s work at Bletchley Park became part of Britain’s extraordinary codebreaking effort.
Other pioneers followed different paths. John von Neumann helped establish ideas underlying generations of stored-program computers. Britain’s Manchester Baby demonstrated stored-program computing in 1948, while machines including EDSAC at Cambridge showed that electronic computers could become practical scientific instruments. Computers then began an astonishing process of shrinking while simultaneously becoming more powerful. Vacuum tubes gave way to transistors, transistors became integrated circuits, and entire processors were eventually placed on single chips. Somewhere along that road appeared a man called Seymour Cray.
Seymour Cray and the computer you could sit on
Few computers have ever possessed the visual personality of the Cray-1. Introduced in 1976, it looked less like an office machine than a piece of futuristic furniture. Its tall cabinets curved into a C-shape and were surrounded at floor level by padded seating. You really could sit on a Cray. The seating was not merely decorative, with parts of the power and cooling infrastructure housed beneath it, while the machine’s distinctive curved design helped keep its internal wiring short. Seymour Cray understood that electrical signals take time to travel, and when engineers are chasing billionths of a second, even the length of a wire matters.
The Cray-1 was capable of roughly 160 million floating-point operations per second at peak performance, an extraordinary figure for its period. It cost millions of dollars and machines of its class were used for demanding scientific and engineering problems, including weather modelling, aircraft research and other forms of advanced numerical simulation.
That word, FLOPS, remains central to supercomputing. It stands for “floating-point operations per second”, essentially a measure of how quickly a computer can perform a particular kind of mathematical calculation involving very large, very small or fractional numbers. A megaflop represents millions of operations per second, a gigaflop billions, a teraflop trillions, a petaflop quadrillions and an exaflop a billion billion. The numbers have become almost comical, but they describe the extraordinary amount of mathematics required by modern scientific computing.
When one processor was no longer enough
For years, one approach to making a computer faster was simply to build an extraordinarily fast central processor, but there are physical limits to how far that can be taken. The answer increasingly became parallel computing. Instead of asking one extraordinarily clever worker to perform a billion jobs consecutively, the work could be divided among thousands, and eventually millions, of computational workers operating simultaneously.
Modern supercomputers are therefore less like solitary mathematical geniuses and more like perfectly choreographed cities. Processors calculate, memory stores information, high-speed networks move enormous quantities of data between different parts of the machine and specialised accelerators perform particular calculations extraordinarily quickly. The difficult part is making the entire orchestra play together.
The mathematics co-processor and the rise of the GPU
Older PC users may remember another specialised piece of hardware: the mathematics co-processor. Early microprocessors were not particularly efficient at some forms of floating-point mathematics, so a separate chip could be installed to perform those operations more quickly. Intel’s 8087, for example, accompanied the 8086 and 8088 generation of processors. Eventually, floating-point hardware became integrated into mainstream CPUs and the separate mathematics co-processor largely disappeared.
Then another specialised processor unexpectedly became one of the engines of modern computing. The GPU, or graphics processing unit, developed primarily to calculate computer graphics. Producing millions of pixels requires performing many similar calculations simultaneously, and scientists and engineers realised that precisely the same talent could be turned loose on scientific mathematics. GPUs subsequently became crucial to modern supercomputers and artificial intelligence. Technology developed to draw computer graphics and games became one of the great numerical workhorses of the 21st century.
Jaguar, Titan, Summit and Frontier
The history of the Oak Ridge National Laboratory in Tennessee illustrates just how quickly this evolution occurred. Its Jaguar supercomputer began operating in the mid-2000s and, through major upgrades, eventually reached petascale computing and topped the TOP500 ranking in 2009. Much of its infrastructure was subsequently transformed into Titan, which combined conventional processors with graphics processors. Titan was followed at Oak Ridge by Summit and then Frontier.
Frontier crossed an historic boundary in 2022 when it became the first system on the TOP500 list to demonstrate exascale performance, exceeding one quintillion floating-point calculations per second on the standard benchmark. That is a billion billion calculations every second. It represented one of those apparently arbitrary numerical milestones that nevertheless signalled a genuine engineering achievement: humanity had constructed a machine demonstrably capable of sustained computation on the exascale.
The fastest computers on Earth
The summit continues to change as new systems arrive. By 2026, machines including LineShine in Shenzhen, El Capitan at Lawrence Livermore National Laboratory, Frontier at Oak Ridge and Aurora at Argonne represented the extraordinary scale of contemporary high-performance computing, while Europe had also entered the exascale era with the JUPITER system in Germany. These are no longer simply fast processors placed in large rooms. They are immense computational installations containing enormous numbers of processing elements, specialised accelerators, memory systems, high-speed interconnections and sophisticated cooling infrastructure.
The Cray name has not entirely vanished into computing history either. Cray eventually became part of Hewlett Packard Enterprise, and the technological descendants of those extraordinary machines can be found among some of today’s most powerful systems. Seymour Cray would probably recognise the obsession immediately, even if the scale would astonish him.
What does a supercomputer actually do?
The purpose of all this machinery is considerably more important than winning a technological horsepower contest. Scientists cannot conveniently create another Earth to discover what increasing atmospheric carbon dioxide will do to its climate. They cannot repeatedly detonate stars to understand supernovae, crash thousands of real aircraft or cars to investigate every conceivable accident, or manufacture millions of potential medicines merely to discover which molecules might work. They can, however, model them.
Supercomputers divide extraordinarily complicated problems into mathematical pieces and calculate what happens to them over time. They help forecast hurricanes, model climate change, design aircraft, investigate nuclear reactions, develop new materials, study galaxies, simulate molecules, analyse genetics and increasingly train and operate artificial-intelligence systems. Large-scale computation has consequently become one of the essential tools of modern science alongside theory, observation and physical experiment.
The little British processor that conquered the world
There is another branch of this story that began not in an enormous American laboratory but in Cambridge. During the 1980s, Sophie Wilson and Steve Furber at Acorn Computers worked on a new processor architecture. Rather than making its instruction set increasingly complicated, they pursued the principles of RISC, Reduced Instruction Set Computing: use a comparatively simple collection of instructions that a processor can execute efficiently and quickly.
The resulting ARM architecture first appeared in silicon with the ARM1 in 1985, containing only around 25,000 transistors. Its descendants would become some of the most important processors ever created. ARM’s combination of performance and low power consumption proved particularly suited to battery-powered equipment and the architecture spread through mobile phones, tablets, embedded computers and countless other devices.
Then the little architecture born at Acorn found its way into supercomputing. Japan’s Fugaku supercomputer uses Fujitsu A64FX processors based on the Arm instruction-set architecture. A processor family whose roots stretch back to a comparatively small British computer company ultimately found itself helping to power one of the most formidable scientific machines ever constructed. The family tree of computing has some wonderfully unexpected branches.
The computers we never see
Yet most computing that governs our lives is not performed by a glamorous supercomputer. When a traffic signal changes, computers may be monitoring detectors, coordinating junctions and responding to changing traffic conditions. When a mobile telephone call is made, computers authenticate the device, manage its connection to the network, allocate radio resources and route the communication. When somebody taps a bank card against a terminal, an intricate chain of computer systems can determine within seconds whether the transaction should proceed.
Aircraft, railways, power distribution, water systems, hospitals, warehouses and telecommunications all depend upon layers of computing that most of us encounter only when something stops working. The modern world has not merely acquired computers; it has become computational.
Grid computing: when many computers become one
There is another approach to enormous computational problems. Instead of constructing one gigantic machine, many separate computers can be connected and the work divided between them. This is broadly the idea behind grid computing and other forms of distributed computation. Universities and research institutions can pool computing resources, while cloud computing extends related principles commercially, allowing organisations to rent vast quantities of processing power without necessarily owning the physical computers themselves.
Even ordinary personal computers have sometimes contributed spare processing capacity to scientific projects. The distinction between “a computer” and “a network of computers” has consequently become increasingly blurred. A computational machine no longer necessarily has a single box around it, or even exists in one building.
Then came artificial intelligence
Artificial intelligence has changed computing again because many modern AI systems require prodigious quantities of computation. Training a large neural network means repeatedly performing mathematical operations across enormous collections of numerical parameters. GPUs and other specialised accelerators are particularly well suited to this kind of parallel mathematics, causing the worlds of artificial intelligence and high-performance computing increasingly to overlap.
The same technological ingredients used to model weather or molecular interactions can therefore also be employed to train neural networks. That does not mean that a supercomputer is simply an enormous artificial intelligence. A computer provides computational power; AI represents a collection of software techniques and mathematical approaches capable of using that power. The distinction matters. A steam engine is not a railway timetable, however much one enables the other.
And now, the qubit
Beyond conventional supercomputing lies something genuinely different. Ordinary computers ultimately store information using bits represented as 0 or 1, whereas quantum computers use quantum bits, or qubits. Through properties of quantum mechanics including superposition and entanglement, qubits can represent and manipulate information in ways unavailable to ordinary digital bits.
This does not simply make a quantum computer an extremely fast conventional computer. It makes it a fundamentally different kind of calculating machine, and the distinction is important because quantum computers will not automatically perform every task faster. Their potential lies in particular classes of problems, potentially including molecular simulation, materials science, optimisation and aspects of cryptography.
Today’s quantum computers nevertheless remain difficult machines to construct and operate. Quantum states are extraordinarily fragile, and unwanted interaction with the surrounding environment introduces errors. Creating reliable logical qubits through quantum error correction is therefore one of the field’s central challenges. The likely future may involve classical supercomputers and quantum processors working together, rather than one simply replacing the other.
The next great machine may be a partnership
That may ultimately be the most interesting lesson in computing history. New machines seldom simply erase everything that came before them. Mainframes survived the arrival of personal computers. Supercomputers survived the microprocessor. CPUs now work alongside GPUs. ARM processors coexist with x86 systems. Cloud computers complement machines sitting underneath desks, while quantum processors are being developed alongside classical ones. Increasingly, humans also work alongside artificial intelligence.
The computer itself is becoming less a single object than an ecosystem of different kinds of machines, each suited to different problems. Seymour Cray shortened wires because fractions of a second mattered. Sophie Wilson and Steve Furber simplified processor instructions because efficiency mattered. Modern supercomputer engineers connect millions of processing elements because parallelism matters. Quantum physicists are attempting to control individual quantum states because entirely new forms of computation may matter.
Different generations and different machines have nevertheless been pursuing essentially the same question: how can we make a machine solve something that yesterday seemed impossible?
From Turing to tomorrow
Alan Turing could scarcely have imagined a smartphone containing billions of transistors. Seymour Cray’s extraordinary machine, once worth millions of dollars and powerful enough to occupy the summit of computing, has long since been surpassed in raw performance by technology enormously smaller and cheaper. Yet the Cray-1 remains important because it represents something more enduring than a benchmark: human ambition to calculate further, faster and more accurately than before.
Today’s exascale computers can perform more than a billion billion floating-point calculations each second. Tomorrow’s systems will go further, while quantum computers may eventually tackle particular problems conventional machines cannot realistically solve. Behind those spectacular machines will remain the quieter computers: changing traffic lights, routing telephone calls, monitoring patients, controlling aircraft, processing payments and performing countless other tasks before breakfast.
We spent much of the 20th century imagining a future in which enormous electronic brains would arrive and transform civilisation. The curious thing is that they did. We simply became so accustomed to living among them that, most of the time, we stopped noticing.
