Premium

AI chip powered by light promises drastic energy savings, 100x efficiency

The prototype chip uses two sets of tiny Fresnel lenses, created through conventional manufacturing methods.

In testing, the new device outperformed traditional chips, accurately classifying handwritten numbers with nearly 98 per cent precision. (Image for representation: Freepik)In testing, the new device outperformed traditional chips, accurately classifying handwritten numbers with nearly 98 per cent precision. (Image for representation: Freepik)

Researchers have developed a new type of computer chip that recognises images and patterns using light instead of electricity, which is one of the most energy-intensive parts of artificial intelligence (AI) technology.

Using light significantly reduces the power needed for these processes, making it 10 to 100 times more efficient than current circuits performing the same calculations. This could enable higher-performance AI models and help decrease the huge electricity demand straining power networks.

AI systems employ a machine learning process called “convolution” to interpret images, videos, and even language. Currently, convolution operations are slow and require a lot of computational power. However, these new devices perform convolutions much faster and with less energy by using lasers and tiny lenses integrated into circuit boards.

Story continues below this ad

In testing, the new device outperformed traditional chips, accurately classifying handwritten numbers with nearly 98 per cent precision.

The prototype chip uses two sets of tiny Fresnel lenses, created through conventional manufacturing methods. These two-dimensional representations of lighthouse lenses are only a small fraction of a human hair’s width. The lenses utilise laser light that has been transformed on-chip using machine learning data, such as images or other pattern-recognition tasks. The AI process concludes by converting the results back into a digital signal.

Volker J Sorger, the Rhines Endowed Professor of Semiconductor Photonics at the University of Florida, who led the research, said, “Achieving a critical machine learning computation at almost zero energy is a leap forward for future AI systems. This is essential to continue expanding AI capabilities in the years to come.”

“This is the first time that this kind of optical computation has been placed on a chip and used in an AI neural network,” added Hangbo Yang, a co-author of the paper and research associate professor in Sorger’s lab in Florida.

Story continues below this ad

This lens-based convolution method not only saves calculation time but is also more efficient in terms of computation. Moreover, using light instead of electricity offers additional benefits. Sorger’s team developed a device capable of processing multiple data streams simultaneously using lasers of different colours.

Yang explained that different wavelengths or colours of light could pass through the lens at the same time. “That is a major benefit of photonics.”

Sorger, who also holds the position of deputy director for strategic projects at the Florida Semiconductor Institute, predicted that chip-based optics will soon play a significant role in all of the AI chips we use daily. “The next step is optical AI computing.”

Since chip manufacturers like market leader Nvidia already incorporate optical components into their AI systems, integrating convolution lenses could become straightforward.

Latest Comment
Post Comment
Read Comments
Advertisement
Loading Taboola...
Advertisement