The July 2022 issue of IEEE Spectrum is here!

Close bar

Data Centers of the Future Could Send Bits Over Infrared Lasers Instead of Fiber Optic Cables

A prototype in a lab at Penn State can beam data at rates of 10 gigabits per second

3 min read
An experimental setup shows a close-up of an infrared laser as it travels through a lens
Photo: Patrick Mansell/Penn State

Go ahead and rip out the hundreds of fiber optic cables that snake between server racks at the nearest data center. Replace them with infrared lasers mounted to the tops of each server rack. Use these lasers to beam data to photoreceptors mounted to more racks across the room. Then, add tiny moveable mirrors to redirect the beams and reconfigure the whole system in an instant.

That’s the vision of Mohsen Kavehrad, a professor of electrical engineering at Penn State. So far, he has built a prototype of one such data link in his lab. Infrared wavelengths are already commonly used to transmit data within fiber optic cables, but now Kavehrad wants to use them to send data over the air. With his system, he has shown that infrared lasers can deliver data rates of 10 gigabits per second.

Kavehrad has published multiple studies on this approach and presented his research today at the Photonics West conference in San Francisco.

Someday, he hopes such lasers could replace the heavy bundles of fiber optic cables found in modern data centers. “If you visit one of these data centers, it's like a jungle, with fiber going between servers,” he says.

In the demonstration system, Kavehrad used lasers to generate an infrared signal with a wavelength of 1550 nanometers, which is a common wavelength used in fiber optic cables. That signal underwent wavelength division multiplexing, which is a technique that packs more signals with multiple wavelengths onto a single laser beam. Then, he sent the beam through an inexpensive lens.

About 15 meters away, he set up another lens and several photodiode receivers. To make the beam steerable, Kavehrad added tiny mirrors, just 2 millimeters in diameter, powered by MEMS, or microelectromechanical systems. The link is bidirectional, which means both ends can send and receive data.

In addition to the infrared signal, the group also broadcast a TV signal using the same setup. They generated it by feeding the entire 1 gigahertz cable TV band into their multiplexer, so it rode along on the same laser beam as the rest of the data. At the other end, they set up an LED TV to show the working channels.

Depending on how many of these links were installed in a data center, Kavehrad thinks his approach could deliver bandwidth and throughput that is as good or better than the fiber optic cables, routers, and switches used today. He says an infrared system should easily be able to handle terabytes of data, given modern improvements in lasers and photodetectors. Kavehrad also hopes the mirrors will allow operators to more quickly respond to fluctuations in demand, and improve the efficiency of these large operations.

Data centers in the U.S. account for about 2 percent of total electricity consumed in the country. Much of that electricity is spent cooling the 400,000 or so servers they contain. Since data centers are built for maximum demand, roughly 30 percent of these servers are idle at any given time. That means a lot of energy is spent cooling servers that aren’t even running.

Kavehrad thinks infrared lasers could allow operators to more easily reconfigure server racks so that all the servers that need cooling are in one area, rather than spread out all over the data center. It’s not yet clear how much electricity this might save, or whether the cost of installing lasers would outstrip the savings. His prototype cost about $20,000 to build, though he expects equipment costs would quickly drop if major companies showed an interest, and if integrated electronics continues to see advances.

Jonathan Koomey, a consultant who has researched the energy efficiency of data centers, says it’s hard to know if Kavehrad’s idea will catch on with Google or Netflix, but suggests it could find a smaller market somewhere, perhaps with supercomputers. “Even if it's not something that finds broad use, maybe there are some narrow niche applications that could be critically important,” he says.

Before testing infrared, Kavehrad and his collaborators at Stony Brook University and Carnegie Mellon University wanted to see if high-frequency millimeter waves could instead be a replacement for cables. These waves fall between infrared and conventional radio waves on the electromagnetic spectrum. Unfortunately, the millimeter waves in their tests attenuated, or lost strength, when transmitted over just 10 meters, and “the interference was a killer,” he says.

Once they switched to infrared, the team played it safe and purchased an amplifier so they could boost the signal’s strength. But found they didn’t need it. Instead, they actually had to weaken the infrared signal once it arrived at the receiver because it was too strong for their equipment to handle. “If you have to actually attenuate [the signal] to go into receivers, that means you're in very good shape,” Kavehrad says.

Still, there are other issues that Kavehrad needs to work out. Just as a laptop hums when it’s working hard, server racks vibrate as they process and transfer data. Kavehrad’s team is concerned this vibration might affect the accuracy of their lasers. “If you start vibrating light that's really focused, that’s going to cause a severe loss of data,” he says.

Editor’s note: This post was updated on 13 February. 

The Conversation (0)

The Future of Deep Learning Is Photonic

Computing with light could slash the energy needs of neural networks

10 min read
Image of a computer rendering.

This computer rendering depicts the pattern on a photonic chip that the author and his colleagues have devised for performing neural-network calculations using light.

Alexander Sludds

Think of the many tasks to which computers are being applied that in the not-so-distant past required human intuition. Computers routinely identify objects in images, transcribe speech, translate between languages, diagnose medical conditions, play complex games, and drive cars.

The technique that has empowered these stunning developments is called deep learning, a term that refers to mathematical models known as artificial neural networks. Deep learning is a subfield of machine learning, a branch of computer science based on fitting complex models to data.

Keep Reading ↓Show less