NVIDIA GPU Operator documentation: Time-Slicing GPUs in Kubernetes, and the time-slicing vs MIG comparison

Explicit: 'Unlike Multi-Instance GPU (MIG), there is no memory or fault-isolation between replicas.' Each pod can run an unlimited number of processes on the underlying GPU and the GPU simply gives an equal share of time. Time-slicing trades the memory and fault isolation of MIG for serving more users. The capacity knob a serving engineer reaches for first is therefore not an isolation mechanism at all.