Exaflop

Computers do a great deal of calculation with decimal numbers, that is, with values like 3.14 or 0.0072. Every single such calculation – an addition, a multiplication – counts as one computing step. How fast a machine is gets measured by how many of these steps it can complete in a second. An exaflop means: a quintillion steps per second, in other words a one followed by eighteen zeros. For comparison: a good laptop manages roughly a hundred billion steps per second. An exaflop machine is thus around ten million times faster.

Why this number counts as a threshold

Exaflop is not an arbitrary mark, but a rung in a long ladder. Before it came the teraflop (a trillion steps) and the petaflop (a quadrillion). Each rung is a thousand times faster than the one before. The leap from petaflop to exaflop took about fifteen years. The first publicly confirmed exaflop machine was the Frontier system in the USA in 2022.

Such machines are expensive and politically charged. An exaflop system costs several hundred million euros and consumes electricity like a small town – often more than twenty megawatts. States build them anyway, because certain questions remain unanswerable without this computing power. These include fine-resolution climate simulations, the development of new medicines, and the safety of nuclear weapons. Whoever owns exaflop machines can pursue research that remains closed off to others.

Since the AI boom, the term has taken on a second meaning. Large language models require enormous amounts of computation to learn. Companies therefore now state their data centers' capacity in exaflops to impress investors. However, these numbers should be taken with a grain of salt, as the next section shows.

How a machine reaches a quintillion steps

No single chip achieves an exaflop. The performance comes from sheer mass: an exaflop machine consists of tens of thousands of processors and graphics chips, housed in racks and connected via extremely fast cables. Frontier, for instance, uses around 37,000 graphics chips. Each chip computes a small part of the task, and in the end the partial results are merged together.

This only works if the task can be divided up in the first place. A weather simulation can be split up well: each chip gets a piece of the Earth’s atmosphere. Other problems cannot be broken down, because each step builds on the previous one. For those, the fastest supercomputer is of little use. A second problem is data transport: often the chips are waiting for numbers from memory instead of computing.

Also important is which level of precision is meant. Classically, exaflops are measured in double precision, that is, with very precise numbers. AI chips, by contrast, deliberately compute less precisely, because that is sufficient for neural networks and is much faster. In this coarser mode of computation, systems easily reach ten or twenty times that figure. So if a company advertises exaflops, it’s worth asking which kind of number is meant.

Exaflops in headlines and product announcements

The figure is most often encountered in the TOP500 list. This ranking appears twice a year and sorts the world’s fastest known supercomputers. For some years now, American and European systems have topped the list; Germany also operates a system of this class, Jupiter in Jülich. Such lists are also a competition between states.

In the tech industry, the term comes up at every major chip announcement. Manufacturers like Nvidia state the performance of entire server racks in exaflops. For investors, this is a signal of how much computing capacity is currently being built. A common misconception here: more exaflops does not automatically mean better AI. Data quality and software are also decisive.

In everyday life, you never encounter an exaflop directly, but you do encounter its results. The weather forecast for the coming days, crash tests done on a computer instead of a real car, the models behind chatbots – all of this comes from data centers of this magnitude.