FLOPs count the calculation steps needed to train or run an AI model. They are a measure of compute.
Example
A large training run needs so many steps that they are given in powers of ten such as 10^25.
For pros
FLOP (a total) and FLOP/s (per second, a chip’s speed) are often confused. The EU AI Act uses training compute as a presumption: above 10^25 FLOP a general-purpose model counts as having systemic risk (Art. 51). The number says little about quality for your task.
