GPU time-slicing

GPU time-slicing shares one physical GPU between several workloads by letting each take turns on the GPU's engines. NVIDIA ships it in two places. In vGPU software, a time-sliced vGPU gives a VM a fixed share of frame buffer while a scheduler rotates the VMs on the GPU. In Kubernetes, the NVIDIA GPU Operator configures the device plugin to advertise a set number of replicas per GPU and hands each to a pod. Unlike Multi-Instance GPU (MIG), replicas get no memory or fault isolation from each other.

vGPU offers three schedulers, best effort, equal share and fixed share, with best effort the default. On a MIG-backed vGPU, the time-slicing happens inside each GPU instance. On the RTX PRO 6000 Blackwell Server Edition, time-sliced profile DC-96Q gives one VM all 96 GB and DC-3Q fits 32 VMs. MIG-backed DC-4-24Q fits four 24 GB vGPUs in a 4g.96gb+gfx instance. In Kubernetes, a config map with replicas set to 4 makes one GPU look like four, though requesting two replicas does not guarantee double the compute.

Use time-slicing when more users per GPU matter more than isolation. It also works on older GPUs without MIG. Use MIG when tenants need guaranteed memory and predictable latency.

Sources

Last verified October 30, 2026.