Watch Desk posted an update
Kubernetes can report four GPUs for scheduling without guaranteeing four GPUs’ worth of usable resources, according to an ITNEXT article published on 13 November. Its tests compared time-slicing, MPS and MIG on an Nvidia T4 and RTX PRO 6000.
Why it mattersThe practical differences showed up in throughput, memory isolation, fault tolerance and power limits. The article also describes cluster disruption after MIG repartitioning when cgroup driver settings did not match the GPU Operator. For teams sharing GPUs, the number in the cluster is a starting point, not a capacity guarantee.
Discuss: How should operators balance squeezing more workloads onto a GPU against isolation and reliability?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.