All terms
Hardware & Systems
Model FLOPs Utilization
How much of a chip's peak compute a workload actually uses, expressed as a fraction.
Definition
Model FLOPs Utilization (MFU) measures how efficiently a training or inference workload uses an accelerator's available floating-point compute. It is the ratio of the math operations per second (FLOPs) a model actually achieves to the hardware's theoretical peak at the relevant precision, where 1.0 would be perfect. In practice, large-model training in the range of forty to sixty percent is considered good; lower values point to bottlenecks in memory bandwidth, communication, or software.