A Multi-Gigahertz "Photonic Chip" Driven by a Frequency Comb Could Blow Current AI Accelerators Away

An international team of researchers have published a paper detailing a new type of processor designed for the age of artificial intelligence, creating a light-based "photonic chip" which uses a frequency comb to allow for the independent processing of multiple wavelengths of light — considerably boosting performance over an electronic equivalent.

"Light-based processors for speeding up tasks in the field of machine learning enable complex mathematical tasks to be processed at high speeds and throughputs," explains the paper's senior co-author Wolfram Pernice, a professor at Münster University, of the breakthrough. "This is much faster than conventional chips which rely on electronic data transfer, such as graphic cards or specialized hardware like TPUs (Tensor Processing Units)."

The chip is built on earlier work carried out at the Ecole Polytechnique Fédérale de Lausanne by Professor Tobias Kippenberg on a chip-based frequency comb, which acts as a light source and allows the photonic chip to process multiple wavelengths of light in parallel without interference. "Our study is the first to apply frequency combs in the field of artificially neural networks," adds Kippenberg. "The frequency comb provides a variety of optical wavelengths that are processed independently of one another in the same photonic chip."

The result is a chip which can operate in a highly parallel fashion, easily outpacing traditional electronic accelerators for artificial intelligence workloads. "Exploiting light for signal transference enables the processor to perform parallel data processing through wavelength multiplexing, which leads to a higher computing density and many matrix multiplications being carried out in just one timestep," says lead author Johannes Feldmann. "In contrast to traditional electronics, which usually work in the low GHz range, optical modulation speeds can be achieved with speeds up to the 50 to 100 GHz range."

The team's work concentrated on using the chip for convolutional processing, but it claims the same technology could be used in a variety of data-heavy artificial intelligence applications including live video processing, cloud computing, and autonomous vehicle systems.

The paper has been published under closed-access terms in the journal Nature.