GPU Time Billing: Queue Time, Active Time and Reserved Capacity Differ

Gpu time billing units: what to decide

A job waits in a queue for twenty minutes and computes for five. Billing twenty-five minutes may be valid for a reserved-capacity offer but misleading for one advertised as active compute. The unit definition settles the distinction.

Record scheduling, allocation and execution timestamps separately. Include GPU count and relevant resource type. A commercial duration may differ from machine occupancy, but the calculation should be reproducible and aligned with the agreement.

An example with the numbers exposed

A job uses two GPUs actively for ten minutes. Under a GPU-minute definition, that is twenty GPU-minutes, not ten. If the offer bills reserved wall-clock time instead, its quantity needs a different clearly stated rule.

The mistake worth avoiding

Failed startup and teardown periods need explicit treatment. Otherwise small jobs can be dominated by overhead the buyer did not expect to pay for.

A practical checklist

• Define billable start and stop conditions.
• Track resource count and class.
• Specify queue and startup treatment.
• Test cancellation and failed-job cases.