Introducing the NVIDIA L4 vGPU service on the ITS Private Cloud

Introducing the NVIDIA L4 vGPU service on the ITS Private Cloud

We're excited to announce the next generation of virtual GPUs (vGPUs) on the ITS Private Cloud, now powered by the NVIDIA L4. The L4 succeeds the NVIDIA TESLA P6 vGPU service that has supported U of T workloads since 2024, bringing significantly more memory, modern AI/ML acceleration, and a broader range of profile sizes to match a wider variety of workloads.

Key Features and Benefits

  • More memory, more flexibility. The L4 provides up to 24 GB of frame buffer per GPU and supports profiles ranging from 6 GB to 24 GB, allowing you to size the GPU to your workload, from lightweight inference to dedicated full-GPU compute.

  • Both graphics and compute profiles available. L4 profiles are offered in two variants: q profiles (RTX Virtual Workstation) for design, visualization, and mixed graphics-and-compute workloads, and c profiles (Virtual Compute Server) for headless AI/ML training and inference. You can pick the profile that matches your use case.

  • NVIDIA AI Enterprise Licensing included. Licensing for both vWS and vCS is included in the service price. Tokens continue to be provisioned automatically through the VMware guest info interface and activated locally on the VM with a simple command.

  • Modern architecture. The L4 is built on NVIDIA's Ada Lovelace architecture and offers substantial improvements in AI inference, video processing, and rendering performance over the previous P6 generation.

  • Available on FD1 and FD2. L4 GPUs are available on the FD1 and FD2 fault domains, expanding GPU availability beyond the single-cluster footprint of the P6 service.

Available vGPU Profiles

To tailor to various needs, we have made available a range of vGPU profiles as shown below:

Profile

Frame Buffer

Use Case

Profile

Frame Buffer

Use Case

nvidia_l4-6q

6GB

High-end design, simulation, multi-monitor workflows

nvidia_l4-8q

8GB

High-res rendering, simulation, lighter AI inference

nvidia_l4-12q

12GB

Large model inference, scientific computing, hi-fi 3D

nvidia_l4-24q

24GB

Full-GPU dedicated workstation or compute VM

nvidia_l4-6c

6GB

Mid-range compute, NLP inference, data science

nvidia_l4-12c

12GB

Large model inference, memory-intensive compute

nvidia_l4-24c

24GB

Full-GPU dedicated compute — AI training, large-scale inference


Why choose the L4?

The L4 is a strong fit for AI inference, machine learning, data analytics, scientific computing, design and visualization, and 3D rendering, the same broad range of workloads supported by the P6, with significantly more headroom for modern AI workloads that benefit from higher memory and newer tensor cores.

Sunsetting the P6 service

The NVIDIA TESLA P6 vGPU service is being sunset alongside the L4 launch. If you have existing P6 vGPUs attached to virtual machines, you'll need to plan a migration to L4 before the P6 hardware is decommissioned.

Getting started

Allocating and configuring an L4 vGPU is a two-step process, and the existing how-to articles cover both new VMs and existing ones.

How-to configure a Virtual GPU on an existing virtual machine

How-to configure and licence a provisioned vGPU

For a complete walkthrough that covers provisioning a new Ubuntu VM with vGPU from scratch, see

How-to provision an Ubuntu VM with NVIDIA vGPU

To verify GPU compute with a Python ML workload after configuration, see

How-to verify NVIDIA vGPU inference on Ubuntu with an ML workload

University of Toronto - Since 1827