Princeton researchers monitored a 319-GPU university research cluster at 100-millisecond intervals over a 10-day period covering 7,399 jobs, and found heavy underutilisation: 50% of jobs averaged below 5% GPU compute (SM) utilisation, and over 70% used less than 20% of GPU memory

Andrew Sheinberg, Samyak Gupta, and Kai Li measured GPU memory and Streaming Multiprocessor (SM) utilization for 7,399 jobs on a university research GPU cluster. They found that over 70% of jobs used less than 20% of GPU memory, and 50% of jobs averaged below 5% Streaming Multiprocessor (SM) utilization. This information was published on 26 Oct 2026 in arXiv:2610.36329.