GPU VMs for Azure Virtual Desktop: When You Need One and Which Size

By David, 2026-11-09, updated 2026-11-09

You need a GPU on Azure Virtual Desktop only for apps that render with one (CAD, 3D modelling, video editing, GIS) or for users who drive more than 3840 × 2160 of total display resolution. For those users, the current choice is the NVadsA10 v5 series, which splits an NVIDIA A10 into slices from one sixth of a GPU upwards, or NCasT4_v3 with an NVIDIA T4. The older NVv3 and NVv4 series retired on 30 September 2026.

The short version: don't buy GPUs for Office, Teams or browsers. For designers and engineers, start with NV6ads_A10_v5 (one sixth of an A10, 4 GB of GPU memory) for light 3D and NV12ads or NV18ads for heavier models, and pilot with the real application. Install the NVIDIA GRID driver, turn on GPU rendering and hardware encoding by policy, and verify it in Task Manager and Event Viewer. GPU hosts cost several times a D-series host, so schedule and autoscale them hard.

Microsoft figures reflect Microsoft Learn as of November 2026, with links at the end. VM prices are the approximate pay-as-you-go rates our AVD cost calculator uses; check the Azure pricing page for your region.

This post expands the GPU section of our Azure Virtual Desktop sizing guide.

When a GPU is worth it, and when it isn't

By default, AVD sessions are rendered on the CPU, and the remote display is encoded in software. That's fine for most work, and a GPU only earns its cost in specific cases:

If the problem is a slow desktop for ordinary users, a GPU won't fix it: check density and memory first, using AVD users per vCPU and the best VM size for AVD.

Which GPU VM sizes AVD supports

Microsoft's GPU acceleration guide lists the sizes supported as session hosts with GPU-accelerated rendering and encoding. As of November 2026:

Two warnings from the same guide. First, NC, NCv2, NCv3, ND and NDv2 sizes aren't generally appropriate as session hosts: they're built for CUDA compute and don't accelerate most applications or the Windows interface. Second, if you still run NVv3 or NVv4 hosts, plan the move to NVadsA10 v5 now; Microsoft's retired sizes list has the details.

NVadsA10 v5 sizes

The NVadsA10 v5 series is the usual answer for AVD because you can buy a slice of a GPU instead of a whole card:

NCasT4_v3 comes in 4, 8 and 16 vCPU sizes with one T4 each (28, 56 and 110 GB of RAM), and a 64 vCPU size with four. It's a sound choice for single-user workstations, but the A10 slices give finer control over cost for a mixed team.

Each NVIDIA-backed VM comes with a GRID license that supports up to 25 concurrent users, so licensing isn't the limit on a multi-session GPU host. CPU, GPU memory and the application are.

How to size GPU hosts

Microsoft's sizing guidance puts power users at 1 user per vCPU, with a minimum of 6 vCPUs and 56 GB of RAM. For GPU hosts, three numbers decide the size:

1. GPU memory per user. The application vendor's minimum frame buffer is the floor. A user on NV6ads gets 4 GB; three users sharing an NV18ads share 12 GB.
2. vCPU per user. Many CAD tools are single-threaded for modelling and multithreaded for rebuilds, so CPU matters as much as the GPU.
3. Pooled or personal. A personal host gives a designer a predictable slice. A pooled GPU host is cheaper per user, but one heavy render slows everyone on it.

A worked comparison for six designers who work 10 hours a day, 5 days a week (216.5 hours a month):

The pooled design is cheaper and gives each designer 6 vCPUs and the same 4 GB share of GPU memory on average, but they share it. Pilot both with your heaviest real models before you commit: the right answer depends on the application, not the price list.

Left running around the clock, the same NV6ads_A10_v5 costs about $737 a month instead of $219, which is why GPU pools pay back scheduling and autoscaling faster than any other pool. Put your own numbers into the AVD cost calculator.

How to turn GPU acceleration on

A GPU VM doesn't use its GPU for sessions until you configure it.

1. Install the right driver. For NVIDIA sizes, only the GRID driver accelerates applications and the Windows interface; the CUDA driver doesn't. The NVIDIA GPU Driver Extension installs GRID automatically on these sizes, and Microsoft supports only Azure-distributed drivers. Bake the driver into your golden image or deploy it with the extension.
2. Enable GPU rendering. In Group Policy or an Intune settings catalog profile, under Administrative Templates > Windows Components > Remote Desktop Services > Remote Desktop Session Host > Remote Session Environment, enable Use hardware graphics adapters for all Remote Desktop Services sessions.
3. Enable hardware encoding. In the same place, enable Configure H.264/AVC hardware encoding for Remote Desktop connections and Prioritize H.264/AVC 444 Graphics mode for Remote Desktop connections. Enabling the 444 mode without the hardware encoding setting can leave you with CPU-based encoding.
4. Optionally, enable HEVC/H.265. Under the Azure Virtual Desktop node, enable Configure H.265/HEVC hardware encoding for Remote Desktop Connections. Microsoft says HEVC gives 25 to 50% better compression than H.264 at the same quality, but it has prerequisites: Windows 10 or 11 session hosts, a desktop application group (not RemoteApp), NVIDIA GRID 16.2 (537.13) or later, multimedia redirection removed from the host, and a client device with an HEVC-capable GPU, the Microsoft HEVC codec and Windows App 2.0.503.0 or later.
5. Restart the session hosts so the policy takes effect.

How to check it's working

• Task Manager: in a session, open a GPU-heavy app and check the GPU on the Performance tab. On NVIDIA sizes, nvidia-smi shows utilization too.
• Event Viewer: under Applications and Services Logs > Microsoft > Windows > RemoteDesktopServices-RdpCoreCDV > Operational, event 170 with AVC hardware encoder enabled: 1 means hardware encoding is on. Event 162 shows Avc444FullScreenProfile for H.264 or HevcProfile for HEVC.
• Connection information: for HEVC, open the connection information window in the session and look for Codecs Used: HEVC under graphics details.

Before you deploy: quota and regions

GPU sizes need their own vCPU quota, and new subscriptions usually start with little or none. Request it before the project date: How to request an Azure vCPU quota increase walks through it. GPU sizes are also offered in fewer regions than D-series, so confirm the size you want is available where your users and FSLogix storage are: see where Azure regions are located.

Watching GPU use with Atlas AVD

Because GPU hosts are expensive, an idle one is the costliest thing in the estate. Atlas AVD includes a GPU Utilization dashboard widget that charts average and busiest-host GPU use across a tenant's N-series session hosts from the Atlas agent's telemetry, so you can see whether designers are using the slice you bought. The usual autoscaling and scheduling apply to GPU pools as to any other. For engineering and design teams, see AVD for engineering and CAD.

FAQ

Does Azure Virtual Desktop need a GPU?

Not for most users. Office, Teams, browsers and line-of-business apps run well on CPU-only hosts. Use a GPU for CAD, 3D, video editing and GIS, or for users with more than 3840 × 2160 of total display resolution.

What is the best GPU VM for Azure Virtual Desktop?

For most teams, the NVadsA10 v5 series. Its partitioned sizes start at one sixth of an NVIDIA A10 with 4 GB of GPU memory (NV6ads_A10_v5), so you can match the slice to the workload. NCasT4_v3 is the other currently supported option.

Are NVv3 and NVv4 still supported for AVD?

No. Microsoft retired both series on 30 September 2026. Move those hosts to NVadsA10 v5 or NCasT4_v3.

Can several users share one GPU on AVD?

Yes. Windows multi-session hosts share the GPU between sessions, and NVIDIA-backed sizes include a GRID license for up to 25 concurrent users. Share only as far as GPU memory and CPU allow, because one heavy job affects everyone on the host.

Why isn't my AVD session using the GPU?

Usually because the GRID driver isn't installed (the CUDA driver doesn't accelerate sessions) or the Use hardware graphics adapters for all Remote Desktop Services sessions policy isn't enabled. Restart the host after setting the policy, then check Task Manager and event 170 in the RdpCoreCDV log.

Can I use NC-series VMs as AVD session hosts?

Only NCasT4_v3. Microsoft says NC, NCv2, NCv3, ND and NDv2 sizes aren't generally appropriate as session hosts because they're built for CUDA compute and don't accelerate the Windows interface.

Last updated 9 November 2026. Microsoft details are drawn from Microsoft Learn as of November 2026: Enable GPU acceleration for Azure Virtual Desktop, the NVadsA10 v5-series and NCasT4_v3-series size pages, Azure VM sizes retirement and Session host virtual machine sizing guidelines. VM prices are approximate pay-as-you-go rates; check the Azure pricing page for your region.