DSI Cluster Impact

The University of Chicago Data Science Institute operates a shared GPU cluster for research and teaching across the University. This site reports what that machine actually delivers — updated nightly, computed from Slurm accounting records, and published with its methodology attached.

Data freshness unknown

No data published yet This site renders metrics that a collector on the DSI cluster publishes into this repository. Until the first run completes, the figures below are empty. See Methodology for how the numbers are produced.
Peak performance not yetAggregate dense FP16 tensor throughput of every installed GPU.
GPUs installed not yetAcross A40, A100, L40S, H100, and H200 nodes.
GPU-hours this year not yetAllocated to research and coursework since 1 January.
Utilization not yetOf GPU-hours the scheduler could offer. See Methodology for the denominator.
Researchers not yetDistinct people who ran at least one job in the last 12 months.
Labs & courses served not yetNamed research groups, courses, and clinic teams in the last 12 months.
Compute delivered not yetCumulative, across all recorded history.
Cloud cost avoided not yetThis year, at public on-demand list rates. See Value Delivered.

How busy is the cluster?

GPU-hours per day

Allocated, idle, and unavailable GPU-hours, stacked to total installed capacity. Drag the slider to change the range; hover for exact figures.

Utilization and availability

Utilization is allocated GPU-hours divided by the GPU-hours the scheduler could actually offer. Availability is the share of installed capacity that was up and schedulable.

Who uses it?

GPU-hours by research group

Most recent full year. Groups below the anonymity threshold are combined into “Other”; their usage still counts toward every total on this page.

When the cluster is busy

Mean GPU-hours allocated by hour of day and day of week, over the last 90 days.

Every figure on this site is derived from Slurm accounting records and published as open JSON. Individual users are never named. Read the methodology, including what these numbers cannot yet tell you, or download the underlying data.