AI Hardware Processing Takes a Leap from 2D to 3D, Significantly Enhancing Computational Power

In a recently published paper in Nature Photonics, a team of researchers hailing from the University of Oxford and partnering with colleagues from the Universities of Muenster, Heidelberg, and Exeter, reveals their pioneering integrated photonic-electronic hardware. This innovation is designed to handle three-dimensional (3D) data, resulting in a significant enhancement of parallel data processing for various AI applications.

While conventional computer chip processing power typically doubles every 18 months, the demands of modern AI tasks are pushing for a doubling in processing power roughly every 3.5 months. This escalating demand underscores the urgent necessity for novel computing paradigms to meet these ever-increasing requirements.

An alternative approach to traditional electronics is the use of light. Light-based processing enables simultaneous calculations through various wavelengths that represent distinct sets of data. In a significant advancement, the same group of researchers, some of whom were involved in this study, presented pioneering research in the journal Nature in 2021. Their work introduced an integrated photonic processing chip capable of performing matrix-vector multiplication, a crucial task for AI and machine learning applications, at speeds surpassing even the fastest electronic methods. This groundbreaking research led to the establishment of Salience Labs, a photonic AI company that spin-out from the University of Oxford.

Taking their research a step further, the team has expanded the processing capabilities of their photonic matrix-vector multiplier chips by incorporating an additional parallel dimension. This advancement, known as “higher-dimensional” processing, harnesses multiple distinct radio frequencies to encode data, pushing the limits of parallelism far beyond previous achievements.

In a practical trial, the research team deployed their innovative hardware to evaluate the risk of sudden death based on electrocardiograms from patients with heart disease. Remarkably, they concurrently analyzed 100 electrocardiogram signals, achieving a remarkable 93.5% accuracy in identifying the risk of sudden death.

The research team also projected that with a relatively modest scale-up of 6 inputs and 6 outputs, this technique has the potential to surpass current electronic processors. This could result in a staggering 100-fold improvement in energy efficiency and computational density. Looking ahead, the researchers anticipate even greater gains in computing parallelism by harnessing additional properties of light, including polarization and mode multiplexing.

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.

Bowei Dong
First author
The Department of Materials, University of Oxford

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.

Harish Bhaskaran
CO-founder of Salience Labs
Department of Materials, University of Oxford

Resources

1. ^NEWSPAPER University of Oxford News. (2023, October 20). From square to cube: Hardware processing for AI goes 3D, boosting. University of Oxford News. [University of Oxford News]
2. ^JOURNAL Dong, B., Aggarwal, S., Zhou, W., Ali, U. E., Farmakidis, N., Lee, J. S., He, Y., Li, X., Kwong, D., Wright, C. D., Pernice, W. H. P., & Bhaskaran, H. (2023). Higher-dimensional processing using a photonic tensor core with continuous-time data. Nature Photonics. [Nature Photonics]