Photonics Spectra BioPhotonics Vision Spectra Photonics Showcase Photonics Buyers' Guide Photonics Handbook Photonics Dictionary Newsletters Bookstore
Latest News Latest Products Features All Things Photonics Podcast
Marketplace Supplier Search Product Search Career Center
Webinars Photonics Media Virtual Events Industry Events Calendar
White Papers Videos Contribute an Article Suggest a Webinar Submit a Press Release Subscribe Advertise Become a Member


Photonic Chip Drives Scalable Data Transfer

Systems that run artificial intelligence programs, such as large language models, face a “bandwidth bottleneck” that can limit the amount of data they transfer between nodes. Using light to send information between the nodes of data centers and high-performance computers could increase the available bandwidth of these systems, while decreasing their energy consumption.

To surmount the bandwidth constraints and high energy costs that limit the performance and scalability of computing systems, researchers at Columbia Engineering, the Fu Foundation School of Engineering and Applied Science, developed a Kerr comb-driven, silicon photonic chip-based data communication link for data transfer.

The millimeter-scale silicon photonic link uses Kerr frequency combs and wavelength-division multiplexing to take one color of light at the input source and create multiple colors at the output. Instead of using a different laser for each wavelength of light, the photonic link requires just one laser to generate hundreds of distinct wavelengths that can simultaneously transfer independent streams of data.

Photonic transmitter chip mounted on a printed circuit board with electrical and fiber optic connections. Courtesy of Lightwave Research Laboratory/Columbia Engineering.

The use of Kerr frequency combs allows clear signals to be sent through separate wavelengths, enabling the silicon photonic chip to transfer large quantities of data over fiber optic cables connecting nodes that may be separated by more than 1 km. The system is massively scalable to hundreds of wavelength channels, enabling it to support parallel optical interconnects for the energy-efficient, hyperscale data centers of the future.

“We recognized that these devices make ideal sources for optical communications, where one can encode independent information channels on each color of light and propagate them over a single optical fiber,” professor Keren Bergman said.

To build the photonic link, the researchers miniaturized all of the optical components onto chips roughly a few millimeters on each edge for generating light, encoded them with electrical data, and converted the optical data back into an electrical signal at the target node.

The link has photonic circuit architecture that allows each channel to be individually encoded with data, with minimal interference from neighboring channels to prevent signals that are sent in different colors from becoming muddled and difficult for the receiver to convert back into electronic data.

Illustration of the hierarchical structure of a disaggregated data center based on Kerr frequency comb-driven silicon photonic links. Courtesy of Lightwave Research Laboratory/Columbia Engineering.

In experiments, the researchers transmitted 16 Gbit/s per wavelength, for 32 distinct wavelengths of light, demonstrating a total single-fiber bandwidth of 512 Gbit/s, with less than one bit in error out of one trillion transmitted bits of data. The silicon chip transmitting the data measured just 4 × 1 mm, and the chip that received the optical signal and converted it into an electrical signal measured just 3 × 1 mm.

“While we used 32 wavelength channels in the proof-of-principle demonstration, our architecture can be scaled to accommodate over 100 channels, which is well within the reach of standard Kerr comb designs,” researcher Anthony Rizzo said.

The researchers also confirmed experimentally that the power penalty of the generated comb lines was negligible compared to a tunable continuous-wave laser source at the same wavelength, demonstrating that each tone behaves identically to an independent continuous-wave carrier from an array of lasers.

Due to their compact nature, the silicon photonic chips can directly interface with computer electronics chips. This reduces the amount of energy consumed during data transfer between nodes, since the electrical data signals only propagate over millimeters, rather than tens of centimeters. The chips can be fabricated by the same facilities used to make the microelectronics chips found in a standard consumer laptop or cellphone, making volume scaling and real-world deployment more straightforward.

“Our approach is much more compact and energy-efficient than comparable approaches,” Rizzo said. “It is also cheaper and easier to scale since the silicon nitride comb generation chips can be fabricated in standard CMOS foundries used to fabricate microelectronics chips, rather than in expensive, dedicated III-V foundries.”

Photonic integrated chip capable of encoding data on 32 independent frequency channels on a U.S. dime for scale. Courtesy of Lightwave Research Laboratory/Columbia Engineering.

The scalable, Kerr comb-driven, silicon photonic link could allow systems to transfer exponentially more data without using proportionately more energy. It represents a promising, realistic approach for data center interconnects to scale to hundreds of wavelength channels, to enable future multi-Tbit/s, chip-to-chip links to operate at low energy.

“What this work shows is a viable path toward both dramatically reducing the system energy consumption while simultaneously increasing the computing power by orders of magnitude, allowing artificial intelligence applications to continue to grow at an exponential rate with minimal environmental impact,” Bergman said.

The next step for the Columbia Engineering team will be to integrate the photonics with chip-scale driving and control electronics to further miniaturize the system.

The research also demonstrates the merit of Kerr frequency combs as optical interconnect sources, establishing them as a practical way to provide integration between nodes in next-generation data centers and high-performance computers.

The research was published in Nature Photonics (www.doi.org/10.1038/s41566-023-01244-7).

Explore related content from Photonics Media




LATEST NEWS

Terms & Conditions Privacy Policy About Us Contact Us

©2024 Photonics Media