Azure GPU-accelerated virtual machines (N-series)
Azure's GPU-accelerated virtual machines are the N-series VM families: ND for large-scale AI training and inference, NC for compute-intensive AI and HPC, and NV and NG for virtual desktops, graphics and cloud gaming. They use NVIDIA or AMD GPUs and are billed like other Azure VMs.1
Also known as N-series VMs, ND-series, NC-series, NV-series, NG-series, Azure GPU VMs
- Intended for
- AI infrastructure and platform teams, ML engineers and researchers, HPC teams, IT teams running virtual desktops
- Runs on
- Azure public cloud regions (each series in selected regions)
At a glance
- What is it?
- GPU-accelerated VMs are one of the Azure virtual machine size types, offered with single, multiple or fractional GPUs. Learn groups them into four families. ND sizes target deep learning and HPC at scale and include ND GB300-v6 and ND GB200-v6, ND-H200-v5, ND-H100-v5, ND_MI300X_v5 with AMD Instinct GPUs, and A100-based series. NC sizes target compute-intensive AI, HPC and rendering, for example NCads_H100_v5, the confidential NCCads_H100_v5 and NC_RTXPRO6000BSE_v6. NV sizes serve virtual desktops and graphics, and NG sizes, with AMD Radeon PRO GPUs, serve cloud gaming and remote desktops.1
- What does it do?
- These VMs give a workload direct access to GPUs in your Azure subscription. Large ND sizes link the GPUs inside a VM over NVLink or AMD Infinity Fabric and connect VMs over InfiniBand with GPUDirect RDMA, so training and inference jobs can span thousands of GPUs in one virtual machine scale set. NC sizes suit training, batch inference and GPU-accelerated analytics, and NV sizes provide GPU-backed virtual workstations. NVIDIA drivers come from the NVIDIA GPU Driver Extension or a manual CUDA or GRID installation.23
- Who needs it?
- AI platform teams that train or serve their own models and need control of the operating system, drivers and frameworks; HPC groups running simulations; and IT teams hosting graphics-heavy virtual desktops. Microsoft's agent adoption guidance lists GPUs and containers for compliance-sensitive or highly customized workloads that need control of the full technology stack.4
- What does it need?5
- An Azure subscription type that can deploy GPU sizes (some series exclude trial and Visual Studio subscriptions)
- VM-family vCPU quota in the target region
- NVIDIA CUDA or GRID drivers, or AMD RCCL-based tooling for ND_MI300X_v5
- Generation 2 VM images for newer ND sizes such as ND GB300-v6 and ND-H200-v5
- What it is not
- GPU VMs are infrastructure, not a model service: you install drivers, frameworks and serving software and operate them yourself. Capacity is not automatic: you need VM-family vCPU quota, and for example Learn states that NCv3 quota starts at zero in each region and that those sizes are not available to trial or Visual Studio subscriptions. Series also retire: Learn gave 30 September 2026 as the NVv3 retirement date and keeps an end-of-life size list.36
Availability and licensing. Available per series in selected Azure regions; check the size page and region list before planning (Learn, checked 26 Oct 2026). Billing follows Azure VM options: pay as you go, reserved instances, Spot VMs and the Azure savings plan for compute. No prices are quoted here; use the Azure pricing calculator for current rates. Retired series move to the end-of-life size list on Learn.7
The problem it solves
Training and serving large models needs many tightly connected GPUs, fast local storage and low-latency networking that few organizations can build quickly on their own premises. Graphics and simulation work has similar needs at smaller scale.
N-series VMs offer these configurations on demand in Azure regions, from a single T4 GPU for inference to rack-scale NVIDIA GB300 NVL72 systems with InfiniBand, so teams can size infrastructure to the job and pay for the time used.6
How it works
- Pick a family and size: ND for multi-GPU training and inference at scale, NC for compute and batch inference, NV and NG for graphics and VDI. Size names encode the family, vCPU count and features.
- Get quota: request a VM-family vCPU quota increase on the Quotas page in the Azure portal; an approved request also raises the regional quota.
- Install drivers: use the NVIDIA GPU Driver Extension or install CUDA or GRID drivers manually, then check with nvidia-smi.
- Scale out: for ND sizes, VMs in the same virtual machine scale set get InfiniBand connections configured automatically, with GPUDirect RDMA.
- Pay: pay as you go, reserved instances, Spot VMs or the Azure savings plan for compute.379
Diagram as a list
AI models & platform
- Frameworks (PyTorch, TensorFlow, NCCL, RCCL)Training and inference softwareConnects to GPU driver extension or CUDA and GRID drivers
- Azure Machine Learning or AKSOptional managed orchestration on top of GPU capacityConnects to N-series VM (ND, NC, NV, NG)
Developer tools
- GPU driver extension or CUDA and GRID driversMakes the GPUs usable by softwareConnects to NVIDIA or AMD GPUs with NVLink or Infinity Fabric
Cloud & hybrid infrastructure
- Subscription vCPU quotaLimits how many GPU vCPUs you can deploy per family and regionConnects to N-series VM (ND, NC, NV, NG)
- N-series VM (ND, NC, NV, NG)The GPU virtual machine
- NVIDIA or AMD GPUs with NVLink or Infinity FabricAccelerators inside the VM
- InfiniBand with GPUDirect RDMAConnects VMs in a scale set for multi-node jobs
Capabilities
Rack-scale Blackwell (ND GB300-v6)5
Each VM has two NVIDIA Grace CPUs and four Blackwell Ultra GPUs with 288 GB each; 18 VMs form one NVL72 rack, and each GPU has 800 Gbps InfiniBand.
Why it matters: Targets high-throughput inference for reasoning and agentic systems and large training jobs.
Limits: No live migration or memory-preserving updates; Generation 2 VMs only; availability by region and quota.
Scale-out H100 and H200 (ND v5)10
Eight NVIDIA H100 or H200 GPUs per VM, with a dedicated 400 Gbps InfiniBand link per GPU and scaling to thousands of GPUs.
Why it matters: Standard building block for distributed training and large-model inference.
Limits: Gen1 images not supported on ND-H200-v5; cluster performance depends on scale set placement.
AMD option (ND_MI300X_v5)2
Eight AMD Instinct MI300X GPUs with 192 GB each, linked by Infinity Fabric, with InfiniBand between VMs.
Why it matters: Gives a non-NVIDIA option for large-memory AI workloads.
Limits: Software stacks must support AMD ROCm and RCCL.
NC for applied AI and confidential GPU6
NCads_H100_v5 offers up to two H100 NVL GPUs; NCCads_H100_v5 extends a trusted execution environment from the confidential VM to the GPU.
Why it matters: Fits batch inference, ML development and protected data processing.
Limits: Smaller GPU counts per VM than ND sizes.
NV and NG for graphics and VDI13
NV sizes target virtual desktops, single-precision compute and video encoding; NG sizes use AMD Radeon PRO GPUs for cloud gaming and remote desktops.
Why it matters: GPU-backed desktops and visualization without local workstations.
Limits: GRID drivers are supported only on listed series and operating systems.
Driver tooling3
The NVIDIA GPU Driver Extension installs CUDA or GRID drivers; Microsoft redistributes GRID drivers that include vGPU licensing in Azure.
Why it matters: Reduces setup errors on N-series VMs.
Limits: Microsoft and NVIDIA support only the documented installation methods and driver versions.
Purchase options7
Pay as you go, reserved instances, Spot VMs and the Azure savings plan for compute.
Why it matters: Matches cost model to steady training clusters or interruptible jobs.
Limits: Spot capacity can be reclaimed; reservations need a term commitment.
Practical use cases
A research team needs to train or fine-tune a large model across many GPUs.
- Approach
- Deploy ND H100 or H200 VMs in a virtual machine scale set with InfiniBand.
- Role of Azure GPU-accelerated virtual machines (N-series)
- GPU VMs provide the compute and interconnect.
- Data, infrastructure and skills
- GPU quota, distributed training code and storage for data and checkpoints.
- Type of benefit
- Shorter training time
- Caveats
- Large GPU quota can take time to obtain; you operate the software stack.
- First step
- Run a small distributed job on two VMs before scaling.
A company wants to self-host an open model so prompts and data stay in its own environment.
- Approach
- Serve the model on NC or ND VMs, or on GPU node pools in AKS, as Microsoft describes for running your own models.
- Role of Azure GPU-accelerated virtual machines (N-series)
- GPU VMs host the model; the agent itself can run on CPU.
- Data, infrastructure and skills
- Model weights, a serving stack and GPU quota.
- Type of benefit
- Data control
- Caveats
- You are responsible for patching, scaling and securing the serving stack.
- First step
- Benchmark the model on one NC VM with your own traffic pattern.
Sources 4
Designers need GPU workstations for CAD and media work from any location.
- Approach
- Provide NV-series virtual desktops with GRID drivers.
- Role of Azure GPU-accelerated virtual machines (N-series)
- GPU VMs supply the graphics capability.
- Data, infrastructure and skills
- Supported OS and GRID driver version, and a desktop delivery method.
- Type of benefit
- Remote access to graphics
- Caveats
- GRID drivers are supported only on listed series and operating systems.
- First step
- Pilot one NV VM with a heavy design application.
Sources 1
Works with
Complementary tools
- Azure Maia and Azure CobaltMicrosoft describes Maia as part of a heterogeneous AI infrastructure that also includes GPUs.
- Azure Kubernetes Service (AKS)Microsoft recommends AKS for hosting large models on large GPU clusters.
- Azure QuantumMicrosoft positions quantum chemistry workflows alongside Azure HPC and AI-accelerated computing.
- Azure Machine LearningAzure Machine Learning compute clusters and serverless compute offer GPU options for training.
Alternative approaches
- Azure Container AppsMicrosoft suggests Container Apps with serverless GPU for small language models.
- Microsoft Foundry ModelsRunning your own models on GPUs is the alternative to calling Foundry endpoints.
Relationship labels follow Microsoft's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Choose the series
Match GPU type, GPU memory and interconnect to the model size using the ND, NC, NV and NG size pages.
Check: The chosen size has enough GPU memory for your model.
Request quota
Open Quotas in the Azure portal, select Compute and request the VM-family vCPU limit you need.
Check: The quota request is approved for your region.
Deploy the VM
Create the VM with a Generation 2 image in a region that offers the size.
Check: The VM is running.
Install drivers
Add the NVIDIA GPU Driver Extension, or install CUDA or GRID drivers as documented.
Check: nvidia-smi lists the GPUs.
Scale out
For multi-node work on ND sizes, deploy VMs in one virtual machine scale set.
Check: A multi-node test job communicates across VMs.
Pick a purchase option
Compare pay as you go, reserved instances, Spot VMs and the savings plan in the pricing calculator.
Check: Your cost model matches how long and how steadily the GPUs run.
Official resources
Deciding whether you need it?
Three free tools apply Microsoft's own published guidance to your situation, and say when Azure GPU-accelerated virtual machines (N-series) or any new technology is not needed.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- Virtual machine sizes overview (opens in a new tab)
- ND family virtual machine size series (opens in a new tab)
- Azure N-series GPU driver setup for Linux (opens in a new tab)
- Technology plan for AI agents (Cloud Adoption Framework) (opens in a new tab)
- ND GB300-v6 size series (opens in a new tab)
- NC family VM size series (opens in a new tab)
- Azure Virtual Machines product page (opens in a new tab)
- What is Azure Machine Learning? (opens in a new tab)
- Increase VM-family vCPU quotas (opens in a new tab)
- ND-H200-v5 size series (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.