FLOPS

in English: floating-point operations per second

advancedwidespread

FLOPS measures how many floating-point calculations a computer performs in one second – this is the computing power's "horsepower".

In plain terms

Just as an engine's power is measured in horsepower, a computer's raw calculation speed is given in FLOPS. The number shows how many floating-point operations—additions or multiplications with decimals—the machine can handle per second.

Artificial intelligence consists almost exclusively of such operations: training a language model and every single response involves countless multiplications and additions. That is why FLOPS shows how much "muscle power" stands behind a data center, a supercomputer, or an AI chip.

The numbers are huge, so you encounter prefixes: gigaFLOPS (billion), teraFLOPS (trillion), petaFLOPS (million billion), exaFLOPS (billion times billion). A modern phone is roughly in the giga–tera range, while the largest supercomputers are in the exa range.

Like horsepower, FLOPS does not tell the whole story: of two machines with the same FLOPS rating, one may be much faster at a specific task if it has more memory or better software.

How it works

Computers store non-integer numbers in so-called floating-point format, which resembles scientific notation: a few significant digits and an order of magnitude.
A floating-point operation is the addition, subtraction, multiplication, or division of two such numbers; FLOPS counts how many fit into one second.
Real performance is measured with standardized benchmark tasks: the TOP500 supercomputer ranking, for example, measures the speed of solving a large system of equations (LINPACK test).
Manufacturers often quote theoretical peak values that the machine would only achieve under ideal conditions; the measured value is lower.
Value also depends on numerical precision: the same chip performs far more operations with lower-precision (e.g., 16-bit) numbers than with 64-bit ones – AI chips typically report the former.

An everyday example

You read in the news that the Debrecen Komondor supercomputer has a performance of roughly 5 petaFLOPS, or that the world's most powerful machines have already crossed the exaFLOPS threshold. The same unit of measurement is used to advertise graphics cards and AI accelerator chips when the manufacturer states how many teraFLOPS they are capable of.

Who is affected

developerThis lets you estimate roughly what size model you can train or run on specific hardware and how long it takes. → Code

entrepreneur, executiveWhen purchasing AI infrastructure, FLOPS is one metric by which hardware can be compared. Cloud offers, however, mostly specify chip type, quantity, and memory, so these should also be checked.

lawyer, citizenThe EU AI Act ties the presumption not to speed per second but to the total number of floating-point operations used for training: above 10^25 operations, a general-purpose model is presumed to pose systemic risk. → Law

bank customer, investorWhen evaluating chipmakers and data center builders, growth in computing capacity is a key metric. This is also expressed in FLOPS, but for AI data centers, the number of chips and power demand (megawatts, gigawatts) are more commonly reported today. → Money

A common misconception

MythThe more FLOPS, the faster the computer is at every task, and the smarter the AI running on it.

In realityFLOPS only measures raw computing speed; actual performance also depends on memory, data movement, and software. Greater computing capacity enables training larger models, but model quality also depends on the data, architecture, and training method, not just the machine's size.

Limits and risks

The gap between theoretical peak performance and actual measured performance can be large, so always check which one is reported.
FLOPS values measured with different numerical precision cannot be directly compared.
The indicator says nothing about energy consumption; efficiency is measured separately in FLOPS per watt.
In AI tasks, memory size and bandwidth are often a tighter bottleneck than raw computing speed.

Similar, but not the same

FLOPs (floating point operations) – FLOPs with a lowercase 's' is the total number of operations performed (for example, 10^25 to train a model), not the speed in operations per second.
tokens per second – This measures how fast a language model generates a response, which depends on the hardware, the model size, and the software.
Clock Speed (GHz) – The clock speed shows how many "ticks" the processor makes per second, but a single tick can involve a very different number of operations depending on the chip's architecture.

Related concepts

Broader

computing capacity, Performance measurement, AI infrastructure

Narrower

PetaFLOPS, ExaFLOPS, FLOPS per watt

On the edge

GPU, Supercomputer, Data center, AI accelerator chip, Benchmark

A brief history

The unit became widespread in scientific computing because earlier metrics did not reflect well the performance of machines working with floating-point numbers. The TOP500 list has ranked the world's supercomputers by FLOPS since 1993. The first machine to cross the exaFLOPS threshold in the list's official measurement was the American Frontier, which reached the top of the list in 2022.

Where things stand

At the end of 2026, FLOPS is the widely used performance metric for supercomputers and AI chips, with top-end machines operating in the exaFLOPS range. In the AI industry, it is increasingly supplemented by energy-based (megawatt, gigawatt) and token-based metrics.

Status: 2026 November

Also known as

FLOPS · flops · flop/s · floating-point operations per second · floating point operations per second · floating-point operations per second · petaflops · exaflops

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.