A computer chip with oscillating beams of light of different colours bouncing across the surface, to represent both light and radio frequencies encoding data.
Artistic rendering of a photonic chip with both light and radio frequencies encoding data. Credit: B.Dong/University of Oxford.

From square to cube: Hardware processing for AI goes 3D, boosting processing power

In a paper published today in Nature Photonics, researchers from the University of Oxford, along with collaborators from the Universities of Muenster, Heidelberg, and Exeter, report on their development of integrated photonic-electronic hardware capable of processing three-dimensional (3D) data, substantially boosting data processing parallelism for AI tasks.

We previously assumed that using light instead of electronics could increase parallelism only by the use of different wavelengths but then we realised that using radio frequencies to represent data opens up yet another dimension, enabling superfast parallel processing for emerging AI hardware.


Dr Bowei Dong, Department of Materials, University of Oxford.

Conventional computer chip processing efficiency doubles every 18 months, but the processing power required by modern AI tasks is currently doubling around every 3.5 months. This means that new computing paradigms are urgently needed to cope with the rising demand.

One approach is to use light instead of electronics – this allows multiple calculations to be carried out in parallel using different wavelengths to represent different sets of data. In ground-breaking work published in the journal Nature in 2021, many of the same authors demonstrated a form of integrated photonic processing chip that could carry out matrix vector multiplication (a crucial task for AI and machine learning applications) at speeds far outpacing the fastest electronic approaches. This work resulted in the birth of the photonic AI company, Salience Labs, a spin-out from the University of Oxford.

Now the team has gone further by adding an extra parallel dimension to the processing capability of their photonic matrix vector multiplier chips. This “higher-dimensional” processing is enabled by exploiting multiple different radio frequencies to encode the data, propelling parallelism to a level far beyond that previously achieved.

As a test case the team applied their novel hardware to the task of assessing the risk of sudden death from electrocardiograms of heart disease patients. They were able to successfully analyse 100 electrocardiogram signals simultaneously, identifying the risk of sudden death with 93.5% accuracy.

Hardware innovations and computations are underpinned by fundamental advances in device materials and innovative uses of their unique properties. Being at Oxford allows one to take risks with ideas, and to build world-class teams that have a wide perspective across different disciplines, skills, and experience levels.


Professor Harish Bhaskaran, Department of Materials, University of Oxford

The researchers further estimated that even with a moderate scaling of 6 inputs × 6 outputs, this approach could outperform state-of-the-art electronic processors, potentially providing a 100-times enhancement in energy efficiency and compute density. The team anticipates further enhancement in computing parallelism in the future, by exploiting more degrees of freedom of light, such as polarization and mode multiplexing.

First author Dr Bowei Dong at the Department of Materials, University of Oxford, said: ‘I am very grateful for the vibrant and collaborative platform that Oxford has provided, giving me the opportunity and courage to touch the frontiers of advanced AI computing hardware and even push it forward. I feel very excited to see where this breakthrough can lead to.'

Professor Harish Bhaskaran, Department of Materials, University of Oxford and co-founder of Salience Labs, who led the work said: ‘This is an exciting time to be doing research in AI hardware at the fundamental scale, and this work is one example of how what we assumed was a limit can be further surpassed.’

 The study ‘Higher-dimensional processing using a photonic tensor core with continuous-time data’ has been published in Nature Photonics.