ailocalinference

← Glossary

Arithmetic Intensity

The ratio of compute work (FLOPs) to memory movement (bytes)

Fundamentals

In plain English

How many math problems you solve per page of the textbook you turn. If you solve 100 per page, you work fast. If you turn a page for just 1 math problem, you spend all your time turning pages.

Technically

The ratio of floating-point operations to memory access (FLOPs/byte). In transformer inference, prompt prefill achieves >100 FLOPs/byte (compute-bound), while autoregressive token generation achieves <2 FLOPs/byte (memory-bound).

Related terms

Roofline ModelMemory BandwidthDecode

More in Fundamentals

DecodeLocal InferencePerplexity (PPL)PrefillRoofline Model