Time-slicing or MIG when one GPU has to serve several pods

One card, four pods, three of them Pending. The cluster is telling the truth. A GPU without a sharing mode is one device. The fourth pod has nowhere to go.
The problem
People turn on both ideas at once, or they turn on MIG on a card that does not support it. The device plugin then publishes a count nobody requested, and every existing manifest breaks.
How I choose
MIG cuts the card into isolated instances. Memory and compute stay apart. It needs a MIG-capable GPU, and changing the geometry interrupts the workloads on that node. I use it when two teams share a card and one job must not be able to take the other team's memory.
Time-slicing leaves one device and lets several pods run on it. There is no memory wall. One pod can allocate the VRAM and the others fail inside the container, which looks like an application bug. I use it for small bursty models that I already know fit together, and I still set a memory limit in the server itself.
The change
I set one strategy in the ClusterPolicy, not both. For time-slicing I point devicePlugin.config at a ConfigMap with the replica count. For MIG I set mig.strategy and the profile, then I drain the node and wait until allocatable shows the slice name. After that I update the pod requests. I do not leave old nvidia.com/gpu: 1 requests pointed at a node that no longer advertises that name.
The setup note is the operator article.
Commercial Engineering & Platform Infrastructure Advisory
Need hands-on senior engineering to audit, scale, or stabilize your infrastructure?
From bare-metal Kubernetes and immutable Talos Linux to drastically cutting cloud infrastructure bills, resolving PostgreSQL deadlocks, and architecting sovereign AI systems, I work directly with production engineering teams to deliver resilient foundations.