Computing at the speed of light

Researchers at AMOLF, the University of Pennsylvania, and the City University of New York (CUNY) have created nanostructured surfaces that can solve equations using light. This discovery opens exciting new opportunities in the field of analog processing based on optical metasurfaces. AMOLF PhD student Andrea Cordaro and his co-authors will publish their findings on January 12, 2023 in his Nature Nanotechnology.

The ever-increasing worldwide need for efficient computing has led researchers in various fields of research to seek alternatives to the current digital computing paradigm. “The processing speed and energy efficiency of standard electronic devices are the limiting factors for new and disruptive applications coming into our daily lives, such as artificial intelligence, machine learning and computer vision,” says Andrea Cordaro. increase. “In this context, analog computing is re-emerging and gaining a lot of attention as a means of complementing traditional architectures.”

Computing at the speed of light

Optical analog processing refers to the use of light to perform analog computations, as opposed to traditional electronic methods that use electricity. One of the main advantages of using light to perform certain computing tasks is that the calculations are performed at the speed that light travels across very thin nanostructured surfaces called metasurfaces, making electronic methods more convenient. It can run much faster than Additionally, optical analog processing is more energy efficient than electronic methods because it does not generate heat like electronic circuits. This makes it suitable for use in high-performance computing applications where speed and energy efficiency are critical.

“For example, in self-driving cars, image detection and processing takes a lot of computational time,” Cordaro says.

“In a previous paper, we showed that optical metasurfaces can be used for very fast edge detection in input images. Detecting the edges of objects such as cars and people is a first step in image processing in many applications. step, and performing this step optically saves processing time and energy.”

little matrix inverse

Cordaro and his colleagues realized that metasurfaces can also be used to perform other mathematical operations. “One of the most frequently occurring problems in many fields, including engineering, science and economics, is the so-called ‘linear inverse problem’. These usually involve matrix inversion, which is a fairly slow math operation,” he says.

The team of researchers developed a thin dielectric nanostructure called a metagrating and incorporated a semi-transparent mirror in the sample to continuously send the signal back to the nanostructure, each time multiplied by the metagrating scattering matrix.

“We use special optimization techniques to design nanostructured arrays or metagrating unit cells that can perform the desired matrix multiplication,” says Cordaro. “Each mathematical problem requires a specific design for the metagrating, but theoretically we could design a surface with multiple parallel gratings to solve several integral equations in parallel. can be solved by

These results demonstrate the potential to solve complex mathematical problems and common matrix inversions at speeds well in excess of those of typical digital computing techniques. In fact, the solution converges in about 349 fs (i.e., less than ten millionths of a second), orders of magnitude faster than conventional processor clock speeds.
Cordaro: “We have demonstrated a powerful new link between nanotechnology and analog computing, which may pave the way for hybrid optical-electronic computing circuits. Our ideas Further development of , will enable us to solve problems of increasing complexity with speed and efficiency previously unthinkable.”