GPU cluster utilization

The share of useful work extracted from installed accelerators, including the time processors sit idle waiting for data to arrive over the network. It is the metric operators use to decide whether to buy more hardware or drive more throughput from what they have.