How can teams share a cluster?
Sharing a cluster and sharing one GPU are different choices. Decide what users need before choosing the mechanism.
Three ways to share
One job receives a GPU; server dependencies may still be shared
Jobs share time; memory and fault isolation differ from MIG
Defined hardware partitions; profiles depend on the product
Three ways to allocate
A job can receive a whole GPU. Multiple jobs can share execution time. Supported GPUs can also be divided into MIG hardware instances with defined resources.
Time slicing does not provide the same memory and fault isolation as MIG. MIG profiles depend on the device and do not replace all the other controls needed for tenant security.
Test workloads together, not only alone.
Quota is not a promise of availability
A quota limits what a team may consume. It does not necessarily reserve that capacity.
If teams require guaranteed access, define a reservation or service policy separately. Also decide which jobs may be interrupted and how they recover.
Try it yourself
Two teams each have an eight-GPU quota in a cluster containing eight GPUs. Can both run eight-GPU jobs at once?
Check your answer
No. Quotas do not manufacture capacity. Likewise, advertising several time-sharing slots on one device does not multiply its memory or arithmetic resources.