Draft

This documentation is in draft and under active review. Figures for the Fall 2026 GPU reservation system are not yet published, and some pages describe behaviour that has not been verified against a primary source. Check with us before relying on anything here.

GPU Access

The Fall 2026 reservation model, documented once. Every audience page links here.


Documented once, and linked from wherever it is needed. One fact, one anchor: a figure documented here is not restated on an audience page.

Every page in this directory is an initial draft. Each opens with a note naming what its writer could not settle — a decision needed, a claim we could not verify, or a figure that does not exist yet. Please read those before treating any page as final.

Page Covers
GPU Classes & Getting One Attached Five size classes, choosing between them, requesting one, and confirming what was allocated.
Reservations Booking a window, the claim window, length caps, best-effort, team mode — and the on-demand launch that creates a reservation for you.
Service Units & Budgets What GPU time costs, what a cancellation costs, and what happens when a budget runs out.
Quotas, Cohorts & Availability What a group is guaranteed, borrowing past it, why availability reads zero, and the status page.
What Ends a Session Idle culling, the end of a window, overstay, preemption and maintenance. None of them is a crash.

← Documentation index