The exponential growth of data traffic in our digital age poses some challenges in processing power. And with automatic learning and AI, for example, in self-driving vehicles and voice recognition, the upward trend will continue. All of this carries a heavy burden on the ability of today’s computer processors to sustain demand.
Now, an international team of scientists has turned on the light to tackle the problem. Researchers have developed a new approach and architecture that combines processing and data storage in a single chip, using light-based or “photonic” processors, showing that they outperform conventional electronic chips by processing information much faster and in parallel.
Scientists have developed a hardware accelerator for so-called matrix vector multiplications, which are the backbone of neural networks (algorithms that simulate the human brain), which are themselves used for machine learning algorithms. Since different light wavelengths (colors) do not interfere with each other, researchers can use multiple light wavelengths for parallel calculations. But to do so, they used another innovative technology, developed at EPFL, as a chip-based “frequency comb” as a light source.
“Our study is the first to apply frequency combs in the field of artificial neural networks,” says Professor Tobias Kippenberg at EPFL, one of the heads of the research. Professor Kippenberg’s research has been a pioneer in the development of frequency combs. “The frequency comb provides multiple optical wavelengths that are processed independently of each other on the same photonic chip.”
“Light-based processors to accelerate tasks in the field of machine learning allow complex mathematical tasks to be processed at high speed and performance,” says co-author Wolfram Pernice, one of the professors who led the research at the University of Münster. “This is much faster than conventional chips based on electronic data transfer, such as graphics cards or specialized hardware such as the TPU (Tensor Processing Unit).”
After designing and manufacturing the photon chips, the researchers tested them on a neural network that recognizes handwritten numbers. Inspired by biology, these networks are a concept in the field of machine learning and are mainly used to process image or audio data. “The conversion operation between input data and one or more filters – for example, can identify edges in an image, are very suitable for our array architecture,” says Johannes Feldmann, now University of Oxford Material Department. Nathan Youngblood (Oxford University) added: “Exploiting wavelength multiplexing results in higher data rates and computational densities, which are operations in every area of the processor that have not been achieved before.”
“This work is a true showcase for European collaborative research,” says David Wright, of the University of Exeter, which runs the EU-funded FunComp project. “Although each of the research teams involved is a world leader in its own way, they were gathering all those parts that really made that work possible.”
The research was published in Nature This week has a wide range of applications: simultaneous artificial data processing (and energy saving) in artificial intelligence, larger neural networks for more accurate predictions and data analysis, large amounts of clinical data for diagnostics, improved evaluation of self-driving vehicle sensor data, and cloud computing infrastructure expanding more storage space, computing power and application software.
Reference: J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. Le Gallo, X. Fu, A. Lukashchuk, “Parallel processing through the nucleus of the integrated photonic voltage”. , AS Raja, J. Liu, CD Wright, A. Sebastian, TJ Kippenberg, WHP Pernice and H. Bhaskaran, January 6, 2021, Nature.
DOI: 10.1038 / s41586-020-03070-1