Usage and quota

What you are holding now, what you have spent, and where the GPU-hours went.

Quota, GPU-hours over time, and cost attribution by project.

What you see

Region	What it shows
Concurrent GPUs / jobs	What you are holding right now against your limits
GPU-hours today	Today’s consumption, against a daily budget if one is set
Storage	Bytes under your run directories, against your storage quota
Cost attribution	GPU-hours grouped by project or by user over a time range

How the hours are counted

Attribution is counted at job end, and a running job contributes its elapsed portion so far. That is why the numbers move during the day and settle overnight, and why a long run appears to grow rather than landing all at once.

The billing clock starts when training actually begins — the running lifecycle mark — not when the executor reports the container as up. The gap between the two is building a virtualenv and pulling weights, which can be several minutes and is not training. Charging for it would systematically overcount, and would penalise exactly the jobs on the slowest network links.

What you can do

• Switch attribution between by project and by user.
• Change the time range.
• Follow a project through to its project page.

When you are refused despite free GPUs

Three different limits refuse for three different reasons, and the message says which:

Limit	Meaning
Concurrent GPUs	You are already holding your maximum. Finish or stop something
Storage quota	Your run directories are over budget. sf clean an experiment you have exported
Scheduling window	This GPU series is outside its permitted hours right now

None of these is the cluster being full. A cluster-full refusal says so, and the scheduler will eventually reserve capacity for a job that has been starved too long.