Phone
+49 4105 770000Call us anytime
Trusted Shops seal
Trusted ShopsGuarantee
ISO 9001 certification
Certified QualityISO9001-2015 certified company
Worldwide shipping
Worldwide shippingWe deliver anywhere
Contact
ContactLeave a message

NVIDIA MIG: GPU Partitioning for AI and Data Centers

·

Hands holding a smartphone showing the NVIDIA logo on a bright screen.

NVIDIA MIG (Multi-Instance GPU) is a technology for hardware-based partitioning of supported NVIDIA GPUs. It enables a single physical GPU to be divided into multiple isolated GPU instances with dedicated compute and memory resources. For IT infrastructure architects and data center operators deploying NVIDIA Ampere, Hopper or Blackwell GPUs in multi-tenant environments, NVIDIA MIG offers a fundamentally different approach to GPU utilization than virtual GPU (vGPU) or time-slicing: hardware partitioning creates true fault isolation between workloads while providing predictable performance. This article explains how NVIDIA MIG works at the hardware level, which GPUs support it, when to use it instead of other sharing methods, and what infrastructure decisions it requires.

How Does NVIDIA MIG Partition a GPU?

NVIDIA MIG divides a single physical GPU into multiple GPU instances by partitioning three core resources: Streaming Multiprocessors (SMs), GPU memory, and memory bandwidth. Each MIG instance operates as an independent GPU to the operating system and applications, with its own memory allocation, fault isolation boundary, and scheduling priority. The partitioning happens at the hardware level. On supported GPUs, SMs, memory slices, and memory controllers are assigned to each GPU instance at the hardware level. A workload in one instance cannot access memory from another instance, and a fault in one instance – whether a software error or a hung kernel – does not propagate to others. This hardware isolation distinguishes NVIDIA MIG from software-based GPU sharing methods where all workloads still share the same GPU memory space and fault domain.

Each MIG instance receives:

  • A fixed number of SMs for computation
  • A dedicated memory slice with its own memory controllers
  • A guaranteed portion of L2 cache
  • Dedicated GPU memory bandwidth
  • Independent compute and copy engines

The instance profiles define how resources are divided. Profile names follow the pattern Xg.YYgb, where X indicates the allocated GPU fraction and YY the memory size. For example, 3g.40gb assigns three-sevenths of the GPU compute resources and 40 GB of memory. The number of instances a GPU can host depends on which profiles are active simultaneously – smaller instances allow more workloads, but each receives fewer resources. MIG instances on separate physical GPUs can communicate over networks using GPUDirect RDMA, subject to configuration constraints.

Which NVIDIA GPUs Support MIG?

NVIDIA MIG is available on selected data center GPUs built on Ampere, Hopper and Blackwell architectures, as well as on RTX PRO Blackwell professional GPUs. Not all enterprise GPUs support the feature – it requires specific hardware partitioning logic introduced with these generations.

Supported GPUs:

  • A100 (Ampere): Up to seven instances. Available in 40 GB and 80 GB variants; the 80 GB model enables larger memory profiles per instance.
  • H100 (Hopper): Up to seven instances. Available in PCIe and SXM configurations with 80 GB.
  • H200 (Hopper): Up to seven instances. HBM3e memory increases per-instance capacity to support larger models.
  • B200 (Blackwell): Up to seven instances. NVIDIA documentation confirms MIG support; final profile specifications depend on memory configuration.
  • RTX 6000 Ada Generation (Professional Workstation): Does not support MIG.
  • RTX PRO 6000 Blackwell Server Edition: Up to four instances. Other RTX PRO Blackwell GPUs also support NVIDIA MIG.

Older architectures – V100 (Volta), T4 (Turing), A10 – do not support NVIDIA MIG. These GPUs lack the hardware partitioning logic required for instance isolation. Workload sharing on these platforms requires vGPU, time-slicing or container-based scheduling, none of which provide hardware-level fault boundaries. The number of available profiles differs by model. An A100 80 GB can create seven 1g.10gb instances or one 7g.80gb instance; combinations of different instance profiles are also possible provided the respective profile and placement rules of the GPU are followed. An H100 offers similar flexibility but with higher memory per fraction. Platform memory capacity determines the maximum profile size – a 40 GB A100 cannot create a 4g.40gb instance.

NVIDIA MIG: Instance Profiles and Memory Allocation

Each MIG profile assigns a specific SM count and memory capacity. The profile determines what workloads fit within the instance and how many instances can coexist.

Common profiles on supported GPUs:

ProfileSM FractionMemory (A100 80 GB)Typical Use Case
1g.10gb1/710 GBInference serving, CI/CD testing
2g.20gb2/720 GBMedium inference models, development
3g.40gb3/740 GBTraining small models, large inference
4g.40gb4/740 GBMixed training and inference
7g.80gb7/780 GBFull GPU as one MIG instance

Actual limits depend on GPU model, memory density and configured profiles – an H200 with HBM3e supports higher per-instance memory than an A100 at equivalent profile fractions.

The memory allocation is exclusive. A workload requesting 12 GB cannot run in a 1g.10gb instance even if other instances are idle – NVIDIA MIG does not dynamically redistribute memory between instances. Administrators must match instance profiles to workload memory requirements before deployment. Oversizing an instance wastes resources; undersizing prevents workloads from starting. Profile granularity varies by GPU generation. Ampere GPUs divide cleanly into sevenths; some NVIDIA Hopper GPU configurations support finer divisions depending on the memory subsystem. The RTX PRO 6000 Blackwell Server Edition creates up to four instances, reflecting a different partitioning design for smaller form factors.

NVIDIA MIG vs. vGPU and Time-Slicing: What Are the Differences?

GPU sharing strategies differ in isolation guarantees, performance predictability and operational flexibility. The right choice depends on whether workloads need fault isolation, whether they run simultaneously, and whether performance variability is acceptable.

FeatureMIGvGPUTime-Slicing
Hardware isolationYes – separate memory, SMsDepends on vGPU configuration; MIG-based vGPU possibleNo – shared GPU memory
Fault isolationHardware-based fault isolation between GPU instancesPartial – hypervisor mediatesNo MIG-based hardware fault isolation
QoS guaranteeYes – dedicated resourcesLimited – shared resourcesNo – scheduling dependent
Concurrent executionYes – instances run simultaneouslyYes – VMs run simultaneouslyTime-division of shared GPU resources
ReconfigurationRequires GPU reset and instance recreationDepends on platform and vGPU configurationImmediate – scheduling policy change
Suited forCloud multi-tenancy, strict SLAsVDI, desktop virtualisationDevelopment, bursty workloads

Use NVIDIA MIG when:

  • Running untrusted or independent tenant workloads where one failure must not disrupt others
  • Service-level agreements require guaranteed compute and memory capacity
  • Multiple Kubernetes namespaces or teams need simultaneous GPU access with isolation
  • Inference serving workloads need predictable latency without interference
  • CI/CD pipelines require dedicated GPU resources for testing without contention

Use vGPU when:

  • Virtualising desktop workloads or VDI deployments where live migration matters
  • Managing GPU allocation at the VM level through existing hypervisor tools
  • Workloads tolerate shared memory and occasional scheduling delays
  • Windows or non-containerised applications need GPU acceleration

Use time-slicing when:

  • Development or prototyping environments where isolation is not critical
  • Workload arrival is bursty and concurrent execution is rare
  • Infrastructure lacks MIG-capable GPUs but needs basic GPU sharing
  • Operational simplicity and zero reconfiguration overhead are priorities

MIG provides determinism: a 3g.40gb instance receives exactly its allocated SMs and memory regardless of activity in other instances. Time-slicing provides flexibility: a workload can use the full GPU when others are idle but gets preempted when contention arises. The trade-off is between guaranteed capacity and opportunistic performance.

What Are the Practical Limitations of MIG?

The hardware isolation from NVIDIA MIG introduces constraints that do not exist in full-GPU or time-sliced configurations. These limits affect workload compatibility, operational flexibility and performance for certain use cases.

Static partitioning: Once instances are created, resources cannot move between them. A 2g.20gb instance cannot borrow memory from an idle 1g.10gb instance. If a workload outgrows its instance, the affected instance must be reconfigured. Depending on the available resources and placement rules, this may require further instance changes. This makes NVIDIA MIG unsuitable for workloads with highly variable memory or compute demands.

Not all partition sizes are available: The sum of active instance fractions must not exceed the available GPU resources. Not all profile combinations are permitted. Administrators must plan profile mixes within the supported combinations documented per GPU model.

Cross-instance communication is restricted: MIG instances cannot directly communicate via GPU memory or NVLink within the same physical GPU. This limits NVIDIA MIG’s applicability to tightly coupled parallel training jobs.

Application compatibility varies: Most inference frameworks and single-GPU training workloads run unmodified in MIG instances. Applications that query GPU topology, assume full SM access, or rely on unified memory addressing may encounter issues. Testing on MIG instances before production deployment is required.

Profile granularity differs by model: An A100 offers seven-way partitioning; the RTX PRO 6000 Blackwell Server Edition supports four instances. Workloads designed for seven-way splits on A100 may not map cleanly to four-way splits on smaller GPUs.

No dynamic resizing: Changing an instance profile requires stopping the affected workloads and adjusting the instance configuration accordingly. A GPU reset is not required for every instance change when MIG mode is enabled. This means NVIDIA MIG is not suitable for environments where resource allocation must adapt continuously to shifting demand.

Size NVIDIA MIG: Plan Workloads and Resources

Choosing the right MIG configuration in professional AI servers depends on how many workloads run simultaneously, what isolation they require, and whether their resource demands are known in advance. A deployment serving ten inference models does not automatically need ten instances – if the models never run concurrently, fewer larger instances may be more economical.

The decision balances three factors:

  • Workload concurrency: how many tasks must execute simultaneously without interference
  • Resource footprint per task: memory and compute required per workload
  • Isolation necessity: whether fault boundaries or QoS guarantees are mandatory

A Kubernetes cluster running fifty microservices that each request GPU resources sporadically can benefit from fewer, larger instances. NVIDIA supports, for example, CUDA MPS on MIG instances, so multiple processes can share the resources of an instance. Conversely, a cloud platform hosting thirty distinct tenants with strict SLAs requires one instance per tenant regardless of utilisation. Memory sizing determines instance count. Seven inference models requiring 8 GB each fit into seven 1g.10gb instances on an A100 80 GB. Four training workloads requiring 35 GB each can, for example, be distributed across four 3g.40gb instances. On an A100 80 GB, a maximum of two such instances are possible simultaneously. For four such workloads, at least two A100 80 GB GPUs are therefore required. Where workloads vary unpredictably, full-GPU allocation or vGPU provides more flexibility than NVIDIA MIG. NVIDIA MIG delivers value where the workload mix is stable, isolation is non-negotiable, and the operational cost of reconfiguration is acceptable.

server-hardware.com supports data center operators, GPU hosting providers, cloud service providers, and system integrators in selecting and configuring high-performance GPU servers for multi-tenant AI infrastructures, inference clusters, and mixed-workload environments. Custom-configured GPU servers enable the deployment of different GPU instance profiles using NVIDIA MIG, with CPU, system memory, and network connectivity tailored to the requirements of each workload. Our server experts advise you on selecting suitable MIG-capable NVIDIA GPUs and help you efficiently allocate GPU resources across multiple customers or applications and make them available for commercial hosting and cloud services.

Technical sources:

NVIDIA MIG User Guide

Emilia GaedeOnline Marketing Manager

Do you have questions?

Our experts are happy to assist you with your Server Hardware questions and offer capable solutions for your problems.

Contact us

Our latest articles