09 · Hardware & Infrastructure

MFU

Model FLOPs Utilization

Model FLOPs utilization: achieved model-training throughput divided by the hardware’s theoretical peak for the relevant arithmetic. Lower MFU can reflect communication, memory traffic, bubbles, or other overhead; it does not mean the remaining hardware is literally unused.

Concrete example

A 35% MFU result means the training workload achieved about 35% of the chosen theoretical peak under that measurement.

Why it matters

One of the most underappreciated metrics in AI economics — where real cost savings and engineering value live.