Polyaxon v3 is coming →

Share GPUs across teams

Give each team its own queues and resource limits, with reusable presets for workloads on a shared Kubernetes cluster.

Uses commercial queues, project controls, and organization-managed presets.

Two teams, one GPU cluster

Research runs parameter sweeps; applied ML runs notebooks and short experiments. Each project uses its own queue, with limits set by the platform team.

Research team

Parameter sweeps

Research queue

GPU quota
Up to 4 GPUs
Concurrency
Up to 4 operations
Dispatch priority
Standard

Applied ML team

Notebooks and short experiments

Interactive queue

GPU quota
Up to 2 GPUs
Concurrency
Up to 2 operations
Dispatch priority
Higher

Two queues, one compute cluster

Shared Kubernetes GPU capacity

Kubernetes places each dispatched workload on a compatible node with available resources.

Illustrative limits, not reserved GPUs. Higher queue priority affects dispatch; it does not evict running jobs.

Connect the cluster and define access

Set limits for teams and workflows

Give teams reusable presets

Explain why a job is waiting

When this setup fits