GPU workloads¶

GPU quota is shared

GPU quota is shared across all workspaces on the platform. Quota is only released when a VM is deleted, not when it is stopped. Please delete your GPU VM as soon as your work is complete.

Do you need a GPU?¶

Not all workloads benefit from a GPU. Before requesting a GPU VM, consider whether your workload actually requires one.

GPU recommended:

Training deep learning models (neural networks, transformers, CNNs)
Running inference on large models (LLMs, image classification)
GPU-accelerated visualisation (3D rendering, certain GIS workloads)

GPU not needed — CPU is sufficient:

Standard statistical analysis, regression, survival models, mixed-effects models
R and Python scripting without explicit GPU libraries
Data wrangling, cleaning, and exploration

If you are unsure, start with a CPU VM. If your workload is taking hours and you are using a GPU library (PyTorch, TensorFlow, CUDA), a GPU VM is likely worthwhile.

Installing GPU drivers¶

GPU drivers are not pre-installed on the platform images. You will need to airlock in the appropriate driver installer for your GPU type from a machine outside the TRE, then install it inside your VM.

Secure Boot must be disabled

GPU driver installation requires Secure Boot to be disabled. All OUH VM images already have Secure Boot disabled — do not attempt to enable it on a GPU VM.

See Airlock requests for instructions on how to bring files into your workspace.

T4 GPU (NCasT4_v3 series) — NVIDIA GRID driver¶

The T4 uses NVIDIA GRID drivers, which Microsoft redistributes directly. You do not need a separate GRID licence — it is included.

Current driver version: GRID 17.3

Download from: Microsoft Azure N-series GPU driver setup — Windows or Linux

Steps:

On a machine outside the TRE, download the GRID driver installer from the Microsoft page above
Airlock the installer into your workspace
Run the installer inside your VM
Verify installation with nvidia-smi in a terminal

A10 GPU (NVadsA10_v5 series) — NVIDIA GRID/vGPU driver¶

The A10 also uses NVIDIA GRID drivers redistributed by Microsoft. The A10 driver is a unified driver supporting both graphics and compute workloads. The minimum supported version is vGPU 17.x — always use the latest available.

Current driver version: vGPU 18 (GRID 18.x)

Download from: Microsoft Azure N-series GPU driver setup — Windows or Linux

Steps:

On a machine outside the TRE, download the GRID driver installer from the Microsoft page above
Airlock the installer into your workspace
Run the installer inside your VM
Verify installation with nvidia-smi in a terminal

Keep the A10 driver up to date

Microsoft periodically deprecates older A10 driver versions on Azure hosts. Running an outdated driver can cause the GPU to become unavailable. Always use the latest version.

H100 GPU (NCadsH100_v5 series) — NVIDIA CUDA driver¶

The H100 is a pure compute GPU and uses standard NVIDIA CUDA drivers — not GRID. Download directly from NVIDIA.

Download from: NVIDIA CUDA Toolkit download page

Select your OS, architecture, and distribution on the NVIDIA page to get the correct installer.

Steps:

On a machine outside the TRE, download the CUDA driver installer from the NVIDIA page above
Airlock the installer into your workspace
Run the installer inside your VM
Verify installation with nvidia-smi in a terminal

CUDA Toolkit vs CUDA drivers

The CUDA driver is required to use the GPU. The full CUDA Toolkit (compilers, libraries, development tools) is only needed if you are writing or compiling CUDA code. For running PyTorch or TensorFlow, the driver alone is sufficient.

Verifying GPU availability¶

After installing drivers, confirm the GPU is visible:

nvidia-smi

You should see your GPU listed with its memory and driver version. If nvidia-smi fails, the driver has not installed correctly — contact the TVS SDE team.

Available GPU VM sizes¶

See the VM sizes guide for full details on available GPU VM sizes, pricing, and guidance on when to use each.

Current quota position¶

Series	Status	UK South quota
T4 (NCasT4_v3)	Available	50 vCPUs
A10 (NVadsA10_v5)	Available	350 vCPUs
H100 (NCadsH100_v5)	Available — extremely limited	Essential workloads only