What Is a Supercomputer? Thousands of Ordinary Chips in a Very Cold Room

A supercomputer is a cluster: many separate compute nodes, each a server with processors and memory, joined by a network and presented to a user as a single machine. The distinguishing features are the ones that make massive parallelism workable:

• •An interconnect with latency measured in microseconds and enormous bandwidth, arranged in a topology such as a dragonfly or fat tree so that any node can reach any other in a few hops. This is usually the most expensive and most distinctive component
• •Accelerators, meaning graphics processors or purpose-built chips, which now supply the great majority of the arithmetic on leading systems because they perform thousands of simple operations simultaneously
• •A parallel file system able to deliver terabytes per second, since a simulation that cannot write its results fast enough is stalled
• •Liquid cooling, generally direct to the chip, because air cannot remove tens of kilowatts per rack
• •A job scheduler allocating thousands of nodes to queued jobs, since nobody uses the whole machine interactively
• •Software written with the message passing interface, a standard for coordinating processes that has been the basis of scientific computing since the 1990s

Performance is quoted in floating point operations per second, and the leading systems are now measured in exaflops, meaning a billion billion per second. The reference ranking, the Top500 list published twice a year since 1993, uses a benchmark that solves a large dense system of linear equations, which has the advantages of being well understood and comparable over three decades and the disadvantage of resembling very little real work: it involves regular, predictable computation with a high ratio of arithmetic to data movement, so a machine can score well and perform poorly on applications that move data irregularly. That is why a second benchmark exists, based on sparse solvers with irregular access patterns, on which most machines achieve a small fraction of their headline figure. The honest summary is that the headline number measures a capability rather than a throughput, and centres judge their own machines on how quickly real codes finish.

The applications share a common shape: a physical system too large or too dangerous to experiment on, described by equations that must be solved on a grid or for a very large number of particles. Weather and climate models divide the atmosphere into cells and step forward in time, and the resolution achievable is limited directly by available computing, which is why forecast accuracy has improved steadily for forty years in step with hardware. Nuclear weapons simulation has been a primary driver of national procurement since testing was banned, and several of the largest machines belong to weapons laboratories. Molecular dynamics simulates protein folding and drug binding atom by atom. Computational fluid dynamics designs aircraft, engines and reactors. Astrophysical simulations follow galaxy formation over billions of years. Seismic imaging processes the enormous datasets that locate oil and gas. And machine learning has become a dominant consumer, with training runs for large models now representing a substantial share of the world's high-performance computing capacity, often on dedicated commercial clusters rather than on academic machines.

Three constraints shape the field. Power is the first and hardest: an exascale system draws in the region of twenty to thirty megawatts, the cost of which dominates the operating budget, and the entire design effort of the last fifteen years has been about performance per watt rather than performance alone. The second is Amdahl's law, which observes that the speedup available from parallelism is limited by whatever fraction of a program must run sequentially, so a code that is ninety-five percent parallel can never exceed a twentyfold speedup however many processors are added, and writing code that is parallel enough is much harder than buying more hardware. The third is reliability: with hundreds of thousands of components, something fails every few hours, so large runs must checkpoint their state to disk periodically in order to restart, and the time spent checkpointing is itself a significant overhead at scale.

The field has become a matter of national prestige and industrial policy. The United States, Japan, China and the European Union all run national programmes, and the list of the fastest machines has been led at various times by American systems at Oak Ridge and Livermore, by Japan's Fugaku, and by Chinese machines whose operators stopped submitting results to the international list after export restrictions were imposed, which means the published ranking is now known to be incomplete. Europe operates several systems through a joint undertaking. The economics have shifted underneath all of this, since the largest clusters in the world by raw capacity are now owned by commercial cloud and artificial intelligence companies and are not submitted for ranking at all. What has not changed is the underlying trade: these machines exist because certain questions can only be answered by simulation, and the answer arrives faster if the electricity bill is paid.

A supercomputer is a cluster of ordinary servers made extraordinary by an interconnect with microsecond latency, accelerator chips doing most of the arithmetic, a parallel file system and direct liquid cooling. Performance is ranked by a linear algebra benchmark that flatters regular computation, so real applications achieve far less. They model weather, climate, nuclear weapons, molecules, fluids and galaxies, and they are limited by electrical power, by the sequential fraction of any program, and by the fact that something fails every few hours.