Close
Scalable, GPU-accelerated infrastructure enabling UCL researchers to develop, train, evaluate, and deploy AI and machine learning models.
Unified AI Platform for Research is built on Kubeflow and supports each stage of the machine learning lifecycle through a single web dashboard. It is delivered and maintained by the ARC Unified AI Services Team.
What you can do with the Unified AI Platform for Research
The platform supports the full machine learning lifecycle in one place. Researchers can develop and test code in interactive notebook sessions, run batch jobs and distributed training on multiple GPUs, build reproducible pipelines that move work from data preparation through to deployment, and serve trained models for evaluation and inference. Each project is paired with RDSS so that datasets and results can move easily in and out of the platform.
Full documentation for using the platform's components is available in the Unified AI user guide.
- Interactive development environments
Researchers can launch notebook sessions running JupyterLab, VS Code or RStudio, configured with the GPU, CPU, memory, and disk space they need. These provide a place to develop and test code interactively before scaling it up into larger jobs.
- Distributed model training
The platform supports batch jobs and distributed training on up to 8 GPUs, with multi-node training coming soon. There is built-in support for common frameworks such as PyTorch and DeepSpeed, allowing researchers to train and fine-tune large models efficiently.
- Reproducible pipelines
Researchers can turn Python code or any containerised job into a scalable, repeatable pipeline, for example preparing data, training a model, then testing and deploying it. A Python SDK provides programmatic access to the platform's components.
- Model serving and inference
Trained models can be deployed behind API endpoints for evaluation and prompt engineering, pulling from sources such as Hugging Face, S3, or local cluster storage. This allows researchers to serve and test models without managing the underlying serving infrastructure.
Requirements
To use the platform, researchers need a VPN connection to UCL or to be working within the UCL network. Some basic command line experience is helpful for working with the platform, but researchers who are new to it are welcome to use it too, and the UAI team will do its best to support those who are learning.
Service tiers and access
The platform is offered in two self-service tiers. The Free Tier provides short-term access for exploration, prototyping, and lightweight use. The Paid Tier provides guaranteed, priority access with extended capabilities for funded projects. Both are requested through My Services, and for the Paid Tier, the team will agree requirements with you before they are costed in Worktribe.
For the Paid Tier, you should include an estimated number of resources you think your project will need, such as GPUs, memory, and storage. The team will then get in touch to discuss the request and advise on the amount of resource that can be allocated, before the agreed resource is costed in Worktribe.
Request access on MyServices
Charge for paid tier
Paid Tier charges are listed in Worktribe under the equipment search. Researchers can find the current rates by searching for "Condenser", which returns the per-unit monthly charges for GPU, CPU, memory, and storage at the relevant rate tier such as Charity, FEC, or Industry. Please contact the team before submitting anything in Worktribe so that we can advise on the right resource allocation for your project.
Support and collaboration
Researchers without an AI or machine learning background, or those needing hands-on support beyond standard onboarding, can work with ARC's Data Science and Research Software Engineering teams through a funded collaboration. Training modules are also in development and will be advertised on ARC's training pages.
Collaborations and consultancy
ARC training
Drop-in Session
We have drop-in session on every Tuesday allowing both in-person and online attendance.
In-person: 90 High Holborn Room 1.41
Online: https://teams.microsoft.com/meet/331105843008638?p=8dvWNofLZeIyhu7LmV
Meeting ID: 331 105 843 008 638
Passcode: Bf7GZ6KA
Frequently asked questions
- Which GPUs do your platform support?
A100 and H200.
- Will my data be hosted on external provider storage such as AWS or Google Cloud?
No. All data on the Unified AI Platform is hosted on UCL infrastructure and premises.
- Do I need to specifically prepare my code to run on the Unified AI Platform?
No. Code runs inside containers on the platform and does not need to be modified from your working implementation.
- Does the platform support sensitive data?
Unfortunately not at the moment. Please contact us for further information or advice.
Useful Link
Knowledge article: Unified AI Platform for Research: Free Tier vs Paid Tier Comparison
User guide: https://kubeflow.arc-unified-ai.condenser.arc.ucl.ac.uk/docs/