New, ARIA-backed project aims to unlock radically cheaper AI hardware

Artificial IntelligenceFundingMathematical, Physical and Life SciencesResearchScience

The University of Oxford is to share in a £50 million award from the Advanced Research + Invention Agency (ARIA), to advance research that could unlock artificial intelligence (AI) hardware at 1/1000th the cost. 

By rethinking communication at multiple levels and innovating across many different disciplines, this project aims to revolutionise the design of AI systems. Our goal is not only to reduce the cost of the AI system, but also to make these sustainable and resilient. 

Professor Noa Zilberman, project leader (Department of Engineering Science)

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. It aims to empower teams of scientists to 'pursue breakthroughs at the edge of the possible' that could unlock world-changing capabilities. 

This new research programme, called Scaling Compute, will address a fundamental challenge facing AI technologies. Using current computing algorithms and hardware, it now costs over £100m to train the largest AI models, with this cost growing at an unsustainable rate. This has profound economic and societal implications, affecting who can build AI, what kind of AI gets built, and at what cost. 

With the development of ever-smaller and faster transistors reaching its limits, further progress can only come by breaking out of existing computer paradigms. Scaling Compute aims to take inspiration from natural systems to challenge the underlying building blocks we use to train AI models. By developing an entirely new library of technologies, the programme has the ambitious aim of reducing the hardware costs required to train large AI models by over a thousand-fold. 

A team from the University of Oxford's Department of Engineering Science will lead one of the twelve projects in the programme. The project aims to develop a system where all the computing resources required to train a frontier AI model (involving hundreds of connected accelerators) are contained within a single computer card. The intention is that this will be stronger than the combined shared compute resources currently available in the university. 

If we succeed in addressing "the communication bottleneck" in AI systems, even small companies and university researchers will be able to afford and run training of frontier AI models, enabling rapid innovation across a wide range of applications.

Professor Martin Booth, leader of the photonics stream of the programme, Department of Engineering Science

This will solve a core problem in AI systems today, which is the communication between devices. Training frontier AI systems requires shuffling huge amounts of data, with a significant portion of processing time spent on transferring information between different parts of a system. A central goal is to ensure that these new communication methods are simultaneously scalable, resilient, sustainable, and manufacturable whilst being capable of efficiently processing the huge volumes of data typically involved in modern computing. 

The team of ARIA R&D Creators will be made up of 16 researchers led by Professor Noa Zilberman (project lead), Professor Amro Awad, Professor Martin Booth, Professor Nick McKeown, Professor Dominic O'Brien, and Dr Patrick Salter at Oxford's Department of Engineering Science. Their broad expertise will enable a highly interdisciplinary approach that will combine innovation in optical systems and photonics, computer architecture and memory systems, computer networks and interconnections, and computing systems design. 

The Oxford team is starting an Industry Affiliates Program to facilitate engagement by industrial partners and collaborators in the research, as well as to advise the project direction. This is an ideal opportunity to meet students for collaboration and future recruitment. For more details, please contact Professor Noa Zilberman.

Further information can be found on the project website.