PFLOPS

PFLOPS is a unit of speed, similar to kilometers per hour. But what it measures isn’t speed on the road, rather how many computational steps a computer can perform in one second. Only calculations involving decimal numbers are counted, for example 3.14 times 2.71. Such calculations are more demanding than adding whole numbers, and they make up the bulk of the work in scientific simulations and in artificial intelligence. One PFLOPS corresponds to one quadrillion such calculations per second, that is, a one followed by fifteen zeros. The abbreviation combines “Peta” for this number with “FLOPS” for floating point operations per second.

The currency in which data centers are traded

Modern AI models don’t emerge from clever programming alone, but from sheer computational volume. During training, a system crunches through vast amounts of data for weeks, adjusting billions of internal values along the way. Whoever has more PFLOPS at their disposal reaches the goal faster, or can even attempt a larger model in the first place. That’s why the figure has become a kind of currency: companies and governments state how many PFLOPS their facilities deliver.

The unit also appears in politics. Export restrictions on AI chips work with threshold values for computing performance. A chip may be sold to certain countries if it stays below a defined limit. The EU’s AI regulatory framework likewise ties stricter obligations to a minimum amount of computing operations consumed for training. A technical metric thus becomes a tool of legislation.

Still, it’s important to note: PFLOPS measures potential, not outcome. A very fast machine with poor software delivers less than a slower one with good software. So the figure indicates what is theoretically possible, not what actually results.

How chips add up to a quadrillion calculations

A single processor core can perhaps manage a few billion calculations per second. Getting from there to a quadrillion is only possible through sheer scale. Graphics chips contain thousands of small compute units working in parallel. And a data center contains thousands of such chips, interconnected via high-speed networks. A system’s PFLOPS figure is, at its core, a sum across all these components.

One detail often causes confusion: the number depends on how precisely the calculations are performed. Scientific simulations require decimal numbers with many digits after the point, usually 64 bit. AI training often gets by with 16 or even 8 bit, because small inaccuracies barely matter there. Coarser numbers mean less effort per calculation. The same chip therefore achieves a multiple of its PFLOPS figure when running in the coarser mode.

That’s why figures given without additional context are of little value. Reputable data sheets specify which precision the value applies to. When comparing two systems, this detail must be checked first, otherwise you end up comparing apples to oranges.

Where the figure turns up in headlines

Twice a year, the TOP500 list of the world’s fastest supercomputers is published. The top spots now lie in the range of over a thousand PFLOPS, which is then called exaflops. The European machine Jupiter in Jülich belongs to this class. Such machines are used for climate models, materials research, and drug development.

In business news, you’ll typically encounter the unit in connection with chipmakers and cloud providers. When a company announces a new data center, the computing performance is listed alongside the investment sum and power consumption. Related units are also common: TFLOPS for one-thousandth of a PFLOPS, EFLOPS for a thousand times as much. A good gaming console today reaches about ten to twenty TFLOPS, while a single AI accelerator lies well above that.

A common mistake is confusing PFLOPS with storage capacity. Terabytes and petabytes describe how much data fits. PFLOPS describes how fast calculations are performed. Both often appear in the same sentence, but they mean entirely different things.