Australian scientists have developed a miniature-sized artificial intelligence (AI) chip that can perform computations based on the power of light, on a speed comparable to that of light.

The prototype of nano photonic chip, which uses the power of the light particles (photons) is entirely in-house in the Sydney Nano Hub in the University of Sydney.

According to the researchers, the prototype can be significant in creating more energy-efficient hardware in the field of artificial intelligence because the global demand of artificial intelligence is still increasing and such technology might reduce the total energy footprint of future computer systems.

Conventional computer chips are made using electricity to control information; that is, to move tiny and charged particles (electrons) with wire. This produces heat.

The prototype of nano photonic chip utilizes light. Light is able to pass through electrically non-resistant materials and hence does not produce heat in the same manner as electricity. The calculation is automatically done through the nanostructures as the light traverses the chip prototype.

The nanostructure of the chip occupies tens of micrometres, the thickness of a human hair. The combination of the nanostructures assists in creating a neural network: the artificial neurons that imitate the human brain to recognise and perform calculations.

The prototype is capable of calculations at the picosecond level, trillionths of a second – the duration in which light exits the nanostructure.

According to the researchers, the benefits of photonics use is that it is much faster and occurs at the speed of light. Light is also used to run the technology as opposed to electricity. This is in comparison to the existing data centres that use huge quantities of water and energy to operate them.

Professor Xiaoke Yi of the School of Electrical and Computer Engineering and director of the Photonics Research Group said that they had re-imagined how the photonics can be used to create new energy efficient and ultrafast computer processing chips.

Artificial intelligence becomes more or less limited to the energy consumption. This study carries out neural computation with light, which has been demonstrated to be faster, more energy-saving and can be made significantly smaller AI accelerators.

The study was published in Nature Communications, and it illustrates that AI models could be made into nanoscale photonic structures capable of manipulating light in such a way that the mathematical operations necessary to carry out machine learning could be implemented.

The researchers tested the nanophotonic chip by training it on over 10,000 biomedical images (breast, chest and abdomen MRI scans, etc.) and validated the technology.

The nanophotonic neural network demonstrated an approximation of 90 to 99 percent classification in simulations and experiments.

The technology provides a way forward to sustainable AI infrastructure which can facilitate the increasing needs of computing without the proportional increase in power usage.

Better, faster, stronger AI hardware

The science of light particles control is known as photonics, which is abbreviated to photon-based electronics. It has been applied in driving technology that is utilized in day-to-day lives like lasers, fiber-optic network and in medical imaging.

However, the harnessing of photonics to computer processing has been a relatively recent discovery and there has been a growing acuity as the need to harness AI demands grows.

The prototype demonstrates how intelligence can be incorporated directly in nanoscale photonic structures, according to PhD student Joel Sved who was instrumental in design and implementation of the prototype.

The Photonics Research Group of the University of Sydney has a long history of over 10 years of research on how to push the limits of photonics as well as how to upgrade our technology.

It involves application of photonics in solving problems in wireless communications and high-technology sensing that are able to detect and measure chemical or biological traces in the environment.

After the successful experiment with the prototype of the nanophotonic chip, the team headed by Professor Yi is currently developing the technology to the level of larger-scale photonic neural networks.