Definition
A FLOP is one piece of decimal-number math (an add or multiply); FLOPs count the total math an AI task needs, while FLOPS measure how many a chip does per second.
At a glance
- FLOPs (lowercase s) = total work; FLOPS (capital S) = speed. Like distance versus a car’s top speed.
- One floating-point operation is a single calculation on a decimal number, e.g. 3.2 times 1.7.
- Counts get huge: mega, giga, tera, peta, exa scale them into millions, billions, and beyond.
- More FLOPS usually means faster AI and lower cost per task.
The distinction that trips people up
FLOPs is the fixed quantity of math a model needs[1]. FLOPS means operations per second and measures hardware speed[3]. The car analogy: FLOPs is the distance to drive, FLOPS is the car’s top speed[2]. Work divided by speed gives time and cost.
Why it matters for buyers
Bigger FLOP counts mean more electricity, chip time, and cost. Training GPT-4 took about 2.1 x 10^25 FLOPs and tens of millions of dollars[4]. Vendors quote FLOPS to advertise GPU speed, but that is a peak rating; real delivered performance is typically a fraction of it[5].
Bottom line
FLOPs is the size of the job; FLOPS is the speed of the machine that finishes it.