Concept·Infrastructure·Added 1 day ago

FLOPs

Also known as: floating point operations, floating-point operations per second, FLOPS, compute

Floating Point Operations. A measure of raw computational work, usually used to describe how much compute was spent training a model. When someone says 'training took 10^24 FLOPs,' they're quantifying the scale of the training run. More FLOPs generally means more capable models, but at steep cost.

FLOPs (sometimes written FLOPS when referring to per-second throughput) is the unit used to measure and compare AI compute. In AI discourse, it usually means the total number of floating-point mathematical operations involved in training a model. It's the currency of the scaling-laws conversation: research shows that more FLOPs, more data, and more parameters generally produce better models.

You'll see FLOPs come up in a few ways. Labs describe training runs in terms of FLOPs to signal how much compute they spent ('trained on 10^25 FLOPs'). Hardware specs quote 'teraFLOPS' or 'petaFLOPS' to describe what a GPU or TPU can do per second. Policy discussions cite national FLOPs thresholds to decide which AI models require oversight.

For builders, FLOPs is mostly background knowledge. You don't need to calculate it. But understanding that 'more FLOPs = more expensive training, usually more capable model' helps you read announcements, understand why frontier models cost what they do, and follow conversations about the AI compute race.

This definition is AI-generated and refreshed weekly. It may contain inaccuracies. Use your own judgment, especially for production decisions.

Related terms