How to Share NVIDIA GPUs with MIG, Time-Slicing, and Resource Quotas

This article explains how to share NVIDIA GPUs in Kubernetes using MIG (Multi-Instance GPU), time-slicing, and ResourceQuotas. It clarifies the operational differences between full-GPU allocation, MIG instances with hardware-backed isolation, and time-sliced replicas that share memory and compute. The guide covers when to use each model based on workload isolation needs, latency sensitivity, and resource contention. It also details configuring MIG profiles, scheduling workloads against MIG resources, setting explicit time-sliced GPU resources, and using Kubernetes ResourceQuota to cap GPU consumption per namespace. Practical steps and validation commands are included for production deployment.