Energy-Efficient Method for Transferring Data Over Fiber-Optic Cables
Researchers demonstrated an energy-efficient method for transferring data over fiber-optic cables that connect computing nodes used to train AI. Technology Briefing
In a study published recently in Nature Photonics, Columbia Engineering researchers demonstrated an energy-efficient method for transferring vastly larger quantities of data over the fiber-optic cables that connect the computing nodes used to train AI. Previous research has addressed this challenge using multiple lasers to generate multiple wavelengths of light. However, a new solution requires only a single laser to generate hundreds of distinct wavelengths of light that can simultaneously transfer independent streams of data.
The millimeter-scale system employs a technique called wavelength-division multiplexing, and exploits devices called Kerr frequency combs which take a single color of light at the input and create many new colors of light at the output. The Kerr frequency combs allowed the researchers to send clear signals through separate and precise wavelengths of light, with space in between them.
The researchers 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. This breakthrough allows systems to transfer exponentially more data without using proportionately more energy. The team miniaturized all of the optical components onto chips a few millimeters on each edge. And they devised a novel photonic circuit architecture that allows each channel to be individually encoded with data while having minimal interference with neighboring channels. Therefore, the signals sent in each color of light don’t become muddled and difficult for the receiver to interpret and convert back into electronic data.
This approach is much more compact and energy-efficient than comparable approaches. And 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 specialized foundries. This provides a straightforward pathway to volume scaling and real-world deployment. The compact nature of these chips enables them to directly interface with computer electronics chips, greatly reducing the total energy consumption since the electrical data signals only have to propagate over millimeters of distance rather than tens of centimeters.
This work shows a viable path towards dramatically reducing system energy consumption and simultaneously increasing computing power by orders of magnitude. This should allow artificial intelligence applications to continue to grow at an exponential rate with minimal environmental impact. In experiments, the researchers managed to transmit 16 gigabits per second per wavelength for 32 distinct wavelengths of light.
This delivered a total single-fiber bandwidth of 512 gigabits per second with less than one bit in error out of one trillion transmitted bits of data. These are incredibly high levels of speed and efficiency. The silicon chip transmitting the data measured just 4 mm x 1 mm, while the chip that received the optical signal and converted it into an electrical signal measured just 3 mm x 1 mm. Both of these are smaller than a human fingernail.
This architecture can be scaled to accommodate over 100 channels, which is well within the reach of standard Kerr comb designs. The next step in this research is to integrate this new photonics chip design with chipscale driving-and-control electronics.