MFU (Model FLOPs Utilization)

Also: model flops utilization

MFU is the share of an accelerator’s peak floating-point compute (FLOPS) actually used. It is the compute-bound counterpart to MBU and matters most for training and the prefill phase of inference; token-by-token decode is typically bandwidth-bound instead, where MBU dominates.

← All terms · Definitions maintained by Mintok. Methodology.