NVIDIA GPU Operator - Time-Slicing GPUs in Kubernetes

Fetched 2026-10-04 (version 26.7). Verbatim: 'Unlike Multi-Instance GPU (MIG), there is no memory or fault-isolation between replicas'; 'Internally, GPU time-slicing is used to multiplex workloads from replicas of the same underlying GPU.' And the sentence that matters for capacity planning: 'A request for more than one time-sliced GPU does not guarantee that the pod receives access to a proportional amount of GPU compute power. Each pod can run as many processes on the underlying GPU without a limit. The GPU simply provides an equal share of time to all GPU processes, across all of the pods.' Time-slicing can be applied cluster-wide or per-node via labels (e.g. Tesla-T4 nodes only), and can be combined with MIG time-slicing of MIG instances. 'MIG support was added to Kubernetes in 2020.' Practical consequence for the topology layer: requesting 2 time-sliced replicas does not give 2 replicas of throughput, and a co-tenant pod on the same physical GPU can take the memory that your KV cache was sized against.