New collaboration to transform the performance and sustainability of AI models

An Imperial-led collaboration is set to investigate new ways to measure and boost the performance of powerful AI frontier models.

A collaboration between Imperial, the University of Edinburgh and the University of Cambridge has been awarded £4.5 million from the Advanced Research + Invention Agency (ARIA), to investigate new and improved ways to measure and boost the performance of powerful AI frontier models over the next three years.   

Established by the UK Parliament in January 2023 and sponsored by the Department for Science, Innovation and Technology, ARIA is the UK’s new research funding agency. Imperial’s new collaboration is one of 12 projects under the new £50 million Scaling Compute programme which has the ambitious aim of increasing and opening up new vectors of progress in the field of computing by bringing the cost of AI hardware down by over a thousand-fold.  

Frontier AI models are state-of-the-art machine learning models that represent the pinnacle of humankind’s current computing capabilities. They operate behind all open AI platforms and can generate text, images, audio, and video on demand to millions of users almost instantaneously.  They have enormous potential to boost the economic and skills productivity across all industries and sectors, as well as transforming the way we live and interact with each other.  

But the current environmental and financial costs for designing, training, and running this technology are unsustainable, and the regulatory and application risks of such a rapidly developing technology are complex and hard to predict. 

ARIA R&D Creator and study lead Dr Aaron Zhao from Imperial’s Department of Electrical and Electronic Engineering says: “This ARIA grant will help us to create the first scalable simulation environment to examine what would happen to the performance and energy usage of existing frontier AI models if we changed or adapted different components, in the same way digital twins are used for examining virtually how different medicines or treatments could affect the human body.  

“We are at a tipping point with the use of frontier AI models. We know they have already had a positive impact across industry, and on our everyday lives through tools such as AI assistants, but the vast carbon footprint and level of financial investment they require currently to operate will have a catastrophic impact on our environment and society if they are not significantly reduced. 

This collaboration brings together some of the UK’s leading experts in machine learning and algorithms, computing systems and AI hardware Dr Aaron Zhao Lecturer in Computer Engineering

“This collaboration brings together some of the UK’s leading experts in machine learning and algorithms, computing systems and AI hardware to allow us to tackle this challenge from every angle.  We can effectively explore what would happen if we went right back to the drawing board and redesigned the computing paradigm the model is based on, as well as investigate the impact of small changes like switching GPU architectures.” 

ARIA Programme Director Suraj Bramhavar said: "The forefront of AI computing hardware has grown from a semiconductor design problem to a massive distributed systems problem. Industry has exquisite tools for simulating the behaviour of individual chips, but lacks a similar toolkit for large warehouse-scale systems. 

"This project aims to bridge this divide. The software tools the team is creating will provide the first widely available map of performance bottlenecks at scale. It will become an essential tool for designing future hardware systems, and will allow the AI hardware community to keep up with the breakneck pace of AI algorithms research."

Join the team

There are PhD and post-doc opportunities available within this study. To find out more about current vacancies and to express an interest please email Dr Aaron Zhao at a.zhao@imperial.ac.uk