MFU
The share of a chip's peak arithmetic speed that a training job actually turns into useful model work.
Also: model FLOPs utilisation
Model FLOPs utilisation compares the arithmetic a model needs per step with what the hardware could theoretically do in the same time. The measure was set out in Google's PaLM paper of 2022. Large training runs commonly report MFU between about 30 and 50 per cent; the rest is lost waiting for memory, for network transfers between GPUs, and to restarts. Raising MFU is worth a great deal of money on a large cluster, which is why providers compete on network design, software tuning and reliability rather than GPU count alone.