Multi-Instance GPU (MIG)

MIG partitions a single physical GPU into up to seven independent instances at the hardware level, not through software virtualization. Each instance receives dedicated streaming multiprocessors (SMs), GPU memory slices, and L2 cache partitions. This physical separation ensures that workloads running on different instances cannot interfere with each other—a fault or memory leak in one instance has zero impact on others. Unlike time-slicing or process-level isolation, MIG provides deterministic latency and throughput for each partition, making it suitable for multi-tenant environments where strict SLAs must be enforced.

MIG supports configurable instance profiles that define the exact fraction of GPU resources allocated to each partition. On an NVIDIA A100 40GB GPU, available profiles range from a 1g.5gb slice (1/7th of compute, 5GB memory) to a 7g.40gb full-GPU instance. Administrators can mix and match profiles—for example, combining three 2g.10gb instances with one 1g.5gb instance on a single GPU. This granularity allows infrastructure teams to right-size GPU allocations for diverse workloads, from lightweight inference tasks to demanding training jobs, maximizing utilization without over-provisioning.

Each MIG instance presents as a fully independent CUDA-capable GPU to the host operating system and applications. This means multiple users, containers, or Kubernetes pods can simultaneously execute different workloads on the same physical GPU without any awareness of each other. A data science team can run a training job on a 4g.20gb instance while the production serving layer handles inference requests on two separate 1g.5gb instances—all on a single A100. This concurrency dramatically improves GPU utilization in shared clusters, reducing idle time and lowering the total cost of ownership for on-premises deployments.

MIG enforces strict resource boundaries at the hardware level, providing predictable performance for each instance. Key QoS mechanisms include:

• Dedicated memory bandwidth: Each instance has guaranteed DRAM bandwidth allocation, preventing noisy-neighbor memory contention
• Isolated error handling: ECC errors or page faults in one instance are contained and reported only to that instance
• Independent GPU reset: A single instance can be reset without affecting others, enabling rapid recovery from hung workloads This deterministic behavior is critical for production AI services where latency percentiles must remain stable regardless of co-located workloads.

MIG is available on NVIDIA GPUs based on the Ampere architecture and newer, including:

• NVIDIA A100: Up to 7 MIG instances (40GB and 80GB variants)
• NVIDIA A30: Up to 4 MIG instances (24GB)
• NVIDIA H100: Up to 7 MIG instances with second-generation MIG enhancements
• NVIDIA H200: Extended MIG capabilities with larger HBM3e memory pools The feature is exclusive to data center GPUs and is not available on consumer-grade GeForce or professional RTX/Quadro cards. MIG must be explicitly enabled via nvidia-smi or the GPU Operator before instances can be created.