Evidence

Research clusters leave most of their GPU memory empty.

Our own cluster analysis and two peer-reviewed studies agree: many GPU jobs need a GPU's compute but not its memory. That unused memory is where Cleanroom runs.

Our analysis

62,827 GPU-hours on a real university cluster

We analyzed 62,827 profiled GPU-hours on a U.S. research university's shared GPU cluster (Sep 21 – Oct 21, 2026), comparing each job's peak GPU memory use with its average GPU compute use.

55.7%

of GPU-hours peaked at or below 25% of GPU memory

71.9%

of GPU-hours peaked below 50% of GPU memory

38.8%

of GPU-hours peaked below 10% of GPU memory

21.7%

of GPU-hours kept compute 75–100% busy while using 25% or less of GPU memory

17.6%

of GPU-hours peaked above 75% of GPU memory

Memory vs. compute

The busiest GPUs often leave the most memory free.

The top row shows GPU-hours when compute was 75–100% busy. Nearly half of them (21.7% out of 45.4%) left at least three quarters of memory empty. These jobs need the GPU's compute, but not its memory.

Share of 62,827 profiled GPU-hours by peak GPU memory (columns) and average GPU compute use (rows).
Compute use0–25%25–50%50–75%75–100%
75–100%21.7%5.9%7.2%10.6%
50–75%9.3%3.8%1.1%2.0%
25–50%11.7%2.7%0.9%2.6%
0–25%13.0%3.8%1.3%2.4%

Hover over or tap a cell to see its details. The highlighted cell shows the opportunity: 21.7% of GPU-hours kept compute 75–100% busy while using a quarter of memory or less.

Source: Cleanroom analysis of a U.S. research university's shared GPU cluster, Sep 21 – Oct 21, 2026. 62,827 paired GPU-hours (peak memory, time-averaged compute).
Peer-reviewed studies

Studies of NERSC's Perlmutter supercomputer find the same pattern.

~50%

of GPU-enabled jobs used up to 25% of GPU memory

Li, Michelogiannakis, Cook, Cooray and Chen, “Analyzing Resource Utilization in an HPC System: A Case Study of NERSC's Perlmutter,” ISC High Performance 2023, LNCS vol. 13948, pp. 297–316.

28.6%

average GPU memory use across 118,276 jobs

37.12% of jobs never exceeded 15% of GPU memory (about 37% of node hours), while only 15.74% reached 75% or more.

Sencan, Kulkarni, Coskun and Konate, “Analyzing GPU Utilization in HPC Workloads: Insights from Large-Scale Systems,” PEARC '25 (ACM).

How we measured

Across the cluster's GPU partition, we paired the peak GPU memory of each profiled GPU-hour with its time-averaged GPU compute use. Measurement period: Sep 21 – Oct 21, 2026.

We also have requests for GPU usage data in progress at other universities.

Sources

  1. Cleanroom analysis of a U.S. research university's shared GPU cluster, Sep 21 – Oct 21, 2026. 62,827 paired GPU-hours (peak memory, time-averaged compute).
  2. Li, Michelogiannakis, Cook, Cooray and Chen, “Analyzing Resource Utilization in an HPC System: A Case Study of NERSC's Perlmutter,” ISC High Performance 2023, LNCS vol. 13948, pp. 297–316. (free preprint)
  3. Sencan, Kulkarni, Coskun and Konate, “Analyzing GPU Utilization in HPC Workloads: Insights from Large-Scale Systems,” PEARC '25 (ACM).

Want to see what your own cluster looks like?

We'd like to compare notes with research computing teams on how the GPUs in their clusters are used.