PFLOP

Computers calculate with numbers that have a decimal point — such as 3.14 or 0.0072. A single such calculation, that is, an addition or a multiplication, is called a floating-point operation in technical language. The English abbreviation for this is FLOP. A PFLOP, pronounced petaflop, is a quadrillion of these: a one followed by fifteen zeros. The prefix “peta” stands for exactly this factor, just as “kilo” stands for a thousand. This unit is used to measure how much computational work a chip or an entire data center accomplishes.

Why AI companies count in petaflops

Modern AI models are created by having a program crunch through enormous amounts of data over the course of weeks. This process is called training. The amount of computation involved is the most important figure for comparing effort and cost. This is precisely why petaflops show up in company announcements and technical articles.

The numbers are hard to grasp. A current language model consumes roughly ten to a hundred billion PFLOP during training. A normal laptop can manage perhaps a hundredth of a PFLOP per second. It would therefore need many millions of years. A data center with tens of thousands of specialized chips completes the same work in a few months.

The unit has also become politically relevant. The EU’s AI legal framework and US regulations tie stricter review requirements to a fixed computing threshold. Anyone training a model above this limit must meet additional requirements. In this way, a technical unit of measurement has become a legal boundary line.

Performance per second versus total work

Two meanings must be kept clearly separate. PFLOP alone denotes an amount of work, just as “kilometer” denotes a distance. PFLOP/s, pronounced petaflops per second, denotes a speed — the counterpart to kilometers per hour. Unfortunately, many texts simply write “petaflops” for both. The context usually makes clear which is meant.

Together, both give you the duration. A system delivering 100 PFLOP/s needs about one million seconds, or roughly eleven days, for a task of 100 million PFLOP. In practice, it takes considerably longer. No computer sustains its peak value continuously, because data must be transported and intermediate results synchronized between chips.

A common misconception: manufacturers' figures are not directly comparable. A chip delivers significantly higher numbers when it calculates using more coarsely stored numbers. Such abbreviated number formats require less memory and less time per operation. That’s why any honest figure must include information about which number format is meant.

Where the number shows up in the news

Petaflops are most commonly seen in reports about new graphics chips and data centers. Manufacturers like Nvidia or AMD state the performance of their systems in this unit. The semi-annual ranking of the world’s fastest supercomputers, the TOP500, also uses it. Its leaders now operate in the range of over a thousand PFLOP/s, which is called an exaflop.

For investors, the unit is a revenue indicator. When a corporation announces it will expand its computing capacity by several hundred thousand PFLOP/s, this signals a multi-billion-dollar order to chip manufacturers. Power consumption is also directly tied to this figure. Data centers of this scale consume as much energy as a small city.

In everyday life, you rarely encounter this unit directly. Nevertheless, it is embedded in every AI response you receive. Because every single query to a language model costs a tiny fraction of the computing power that was previously measured — and paid for — in petaflops.