Imagine counting how many brushstrokes a painter can make in one second. FLOPs do the same for a computer: they measure how many floating-point operations (additions and multiplications on decimal numbers) a machine can perform. It is the fundamental unit of computing power in artificial intelligence.
Two meanings, one letter
Beware the trap: that little "s" changes everything.
- FLOPs (operations) = the total work of a computation. We say training a large model cost "10²⁵ FLOPs".
- FLOP/s (per second) = the speed of a chip. A modern GPU reaches tens of PetaFLOP/s (10¹⁵ operations per second).
One is an amount of work, the other a rate — like distinguishing the length of a trip from the speed of the car.
The scale of prefixes
| Prefix | Notation | Value |
|---|---|---|
| Tera | TFLOPs | $10^{12}$ |
| Peta | PFLOPs | $10^{15}$ |
| Exa | EFLOPs | $10^{18}$ |
Why it matters for AI
The training cost of a model can be estimated simply. For a model with $N$ parameters trained on $D$ tokens:
$$C \approx 6 \times N \times D$$
This formula explains why frontier models demand colossal compute budgets: doubling parameters and data nearly quadruples the cost. FLOPs have thus become a strategic currency — so much so that regulations now govern models exceeding certain compute thresholds.
FLOPs don't tell you whether an AI is smart — only how much raw effort went into building it.