Model FLOPs Utilization is an efficiency metric that expresses the ratio of the floating-point operations a model actually needs to the peak FLOPS the hardware could theoretically deliver over the same time. A high MFU means GPUs spend their cycles on useful math rather than waiting on memory or network. Operators track it to judge how well a training cluster's interconnect, software, and parallelism strategy are performing.
Category: IT & Compute Infrastructure · Subcategory: AI & HPC Infrastructure · Part of the AIDC Technical Terminology reference.