UCSD AWS Builder Club GPU pool
A shared GPU pool, live. Anyone can read this page and nothing on it names a person. Updated 25 Aug 2026 17:08 UTC.
demo data generated by running the scheduler against a simulated clock. Nobody here is a real person.
Pool utilization
A break in a line means nothing was running. That is different from a GPU sitting idle, and only one of the two is a problem.
Outcomes
| window | completed | interrupted, resumed | waiting to resume | reclaimed idle | failed |
|---|---|---|---|---|---|
| this week | 7 | 1 | 0 | 1 | 2 |
| last 30 days | 45 | 1 | 0 | 1 | 2 |
A job that lost its machine and carried on is counted once, as completed, and again under interrupted. "Waiting to resume" is only the jobs that are mid-restart at this moment, so it is usually zero and that is the point.
Spend
Reclaimed credit is the part of a reservation an idle job never got to spend. It assumes the job would have kept holding the GPU, which is what it was doing when idle detection caught it.
Where the work ran
| capacity | kind | jobs | gpu-hours | share |
|---|---|---|---|---|
| ec2-spot | cloud | 39 | 93.3 | 72% |
| lab-a6000 | local | 15 | 36.0 | 28% |
Activity
| who | gpu | where | state | ran for | peak gpu |
|---|---|---|---|---|---|
| user-a | a10g | cloud | running | 50m | 85% |
| user-f | a6000 | local | running | 50m | 85% |
| user-c | a100 | cloud | allocating | 35m | - |
| user-b | t4 | cloud | allocating | 35m | - |
| user-e | l4 | cloud | cancelled | 20m | 85% |
| user-c | t4 | cloud | failed | 20m | 85% |
| user-d | a10g | cloud | reclaimed | 55m | 85% |
| user-b | a10g | cloud | failed | 6.2h | 85% |
| user-c | a100 | cloud | completed | 4.9h | 85% |
| user-a | a6000 | local | completed | 2.9h | 85% |
Members appear as stable labels rather than names, assigned by when they first used the broker. What they were running is theirs and is not published.
Running gpu-broker · the same numbers as JSON