Utilization
How much of the cluster’s capacity turns into delivered compute, and how much of it is idle or unavailable.
Data freshness unknown
Utilization (of available)
79.0%Year to date. Excludes hours when hardware was down or in maintenance.
Utilization (of installed)
75.6%Year to date, against every GPU-hour on the floor including downtime.
Availability
95.6%Share of installed GPU-hours that were up and schedulable.
GPU-hours delivered
696.2kYear to date.
Allocated, not measured
These figures describe GPU-hours allocated by the scheduler. Slurm
knows a GPU was assigned to a job; it does not know whether that GPU was
busy. Per-device utilization requires DCGM exporters, which are not yet
deployed — see Methodology.
Nothing on this site claims measured device utilization until they are.
Capacity over time
GPU-hours per day
Stacked to total installed capacity: what was allocated, what sat idle, and what was unavailable.
Utilization and availability rates
Where the hours go
Demand by GPU model is not available
Slurm's accounting on this cluster records that a job used a GPU, but not
which model —
AccountingStorageTRES tracks gres/gpu
with no per-model breakdown, so every GPU-hour arrives unattributed. Jobs
can request a specific model and the scheduler honours it; the type
simply is not retained in the accounting record. Until that changes there is
no honest way to chart demand per generation, so nothing is charted here.
The installed mix is on Capacity.
By partition
By quality of service
The cluster runs three QoS tiers — general,
protected, and interactive. Preemptible work
running in gaps is capacity that would otherwise have been wasted.
Usage by hour and weekday
Mean GPU-hours allocated, last 90 days.