Edit

Supported Azure virtual machine (VM) sizes for AKS-managed GPU node pools

Overview

Azure Kubernetes Service (AKS) can install and maintain GPU software for supported NVIDIA GPU-enabled nodes. Use this article to choose a virtual machine (VM) family, then follow Create an AKS-managed GPU node pool to configure the feature and verify your GPU workload deployment.

Supported GPU-enabled VM sizes

AKS-managed GPU node pools currently support GPU-enabled VM sizes in the NC, ND, and NV families, subject to the exclusions in the following table. The GPU model varies by VM series. Check regional availability before you choose a size.

Azure VM family Supported GPU series Exclusions and retired series
NC family NVIDIA GPU series, including NCasT4, NC A100, and NC H100. Non-NVIDIA GPU series aren't supported.
ND family Include ND A100 v4, NDm A100 v4, and ND H100 v5. ND GB200 v6 and ND GB300 v6 are supported in public preview (see NVIDIA Grace-Blackwell GPU concepts on AKS to learn more). ND MI300X v5 isn't currently supported because it uses AMD GPUs.
NV family NVads A10 v5. NVads V710 v5 isn't currently supported because it uses AMD GPUs. NVv3 and NVv4 retired on September 30, 2026.
NG family None. The NG family isn't currently supported. NGads V620 uses AMD GPUs.
NP family None. The NP family isn't supported. NP-series uses field-programmable gate arrays (FPGAs), not GPUs.

Node pool considerations

  • Use an AKS node pool with --os-type Linux. For Windows GPU workloads, see Use Windows GPUs on AKS.
  • Confirm that your subscription has sufficient quota for the selected VM family in the target region. Regional availability doesn't guarantee deployment capacity.

Note

GPU-enabled VMs contain specialized hardware subject to higher pricing and regional availability. For more information, see the pricing tool and regional availability.

Match VM resources to workload requests and limits

  • Check the specifications of the exact GPU VM size you plan to use, because GPU models, GPU counts, GPU memory, CPUs, storage, and network performance can differ within the same VM family.
  • Choose a VM size with enough GPU memory for the peak memory usage of your GPU workload. Host memory limits don't control GPU memory. In the NVads A10 v5 series, Standard_NV6ads_A10_v5, Standard_NV12ads_A10_v5, and Standard_NV18ads_A10_v5 provide 1/6, 1/3, and 1/2 of a physical NVIDIA A10 GPU, respectively. On these sizes, requesting nvidia.com/gpu: 1 doesn't provide a full physical GPU. Check the selected size's GPU memory allocation rather than inferring it from the Kubernetes resource count.
  • Check the network bandwidth and connections between GPUs that your distributed workload requires. GPU requests don't reserve network bandwidth or enable remote direct memory access (RDMA), so configure any required network software separately.
  • Multi-Instance GPU (MIG) is supported on select NVIDIA GPU VM sizes in AKS. For supported hardware and compatible GPU instance profiles, see Create a multi-instance GPU node pool in AKS. To configure MIG with AKS-managed GPU software, see Create a managed MIG node pool.
  • Test your application with representative model sizes, batch sizes, and numbers of concurrent requests, then adjust resource requests and limits based on observed usage.

Next steps

Create an AKS-managed GPU node pool and run a sample GPU application in your AKS cluster.