GPU Access¶
From Fall 2026, a reservation system manages GPU access on PARASOL (formerly DSMLP). If your course, project, or research needs a GPU, start here to choose a GPU class, reserve time, and plan the Service Units your sessions spend.
| Page | Covers |
|---|---|
| GPU Classes | GPU class sizes, choosing and requesting a class, preset classes on Datahub, a missing or misspelled class label, and confirming the allocation. |
| Workloads by GPU Class | Estimating GPU memory, differences within a class, and example course and research work for each class, with the cards and slices behind it. |
| GPU Hardware & CUDA | Specifications of the cards and MIG slices that can back a GPU class, how MIG slices divide a card, supported CUDA versions, and CUDA profiling. |
| Reservations | Reservation types, booking ahead and launching without a booking, the claim window, length caps, best-effort reservations, team mode, and routes to more time. |
| Service Units & Budgets | What spends Service Units, hourly rates, budget sizes and windows, the cancellation penalty, and the routes available when a budget runs out. |
| Quotas, Cohorts & Availability | GPU quotas, borrowing beyond quota, cohorts, the symptoms of a full cluster, the status page, and requesting a quota increase. |
| What Ends a Session | Idle culling, the end of a reservation window, overstay, preemption, and maintenance closures. |