Chamber supports two workload classes that determine how workloads access GPU capacity: Reserved and Elastic.
Overview
| Class | Capacity Source | Preemptible | Best For |
|---|
| Reserved | Team’s reservation | No | Production training, SLA workloads |
| Elastic | Reserved capacity first, then idle pool | Yes | Experiments, batch processing, dev work |
Reserved Workloads
Reserved workloads use capacity from the team’s reservation. They are guaranteed resources and cannot be preempted.
Use reserved for:
- Long-running training that can’t be interrupted
- Workloads with delivery commitments or deadlines
- ML pipelines that must complete reliably
Elastic Workloads
Elastic workloads use idle capacity. They maximize utilization but can be preempted when reserved workloads need resources.
Use elastic for:
- Experiments and exploratory work
- Development and debugging
- Batch workloads that can checkpoint and resume
Design elastic workloads to checkpoint periodically so they can resume if preempted.
