4 August 2010—The race to replace copper wiring with optics in chip-to-chip communications reached a new milestone last week as Intel announced it had produced a system using silicon-based photonics to transmit data between printed circuit boards at 50 gigabits per second.
”We’re bringing silicon manufacturing to optical communication,” says Mario Paniccia, director of Intel’s Photonics Technology Lab. ”It changes the way in the future that we’re going to connect.” Until recently, optical communications was done using exotic semiconductors and other expensive components. Making such systems in silicon should lower their price and allow for easy integration into computers.
The prototype, which was announced at the Integrated Photonics Research Conference in Monterey, Calif., takes several discrete technologies that Intel has invented over the past few years and combines them into one package. These include a hybrid silicon/indium phosphide laser, a silicon modulator operating at 40 Gb/s, and a germanium detector, also operating at 40 Gb/s. The company has brought those together into a four-channel link, with each channel operating at 12.5 Gb/s, for a total bandwidth of 50 Gb/s. ”We’ve always said the real value of silicon photonics is in integration,” Paniccia says.
But Greg Young, president and CEO of Luxtera, a small silicon photonics company in Carlsbad, Calif., says his company’s 40-Gb/s active optical cable, called Blazar, and its OptoPHY transceiver already provide similar high-rate optical transmission. Luxtera also uses a laser of indium phosphide bonded to silicon and silicon modulators and photodetectors. In June, the company announced that its modulators could handle 30-Gb/s data rates, a step toward a 100-Gb/s interconnect.
”We’re happy they believe in the same thing we believe in, but they’re announcing that they’ve demonstrated something very close to what we’re selling,” Young says. ”It’s as if they think they live in a parallel universe where Luxtera doesn’t exist.”
Young says his company’s near-term competition is with makers of nonsilicon optical interconnects, which use vertical-cavity surface-emitting lasers as their light sources and multimode fiber to carry the light. The problem with that technology, he says, is that it doesn’t easily scale up to higher speeds. That’s because the cost of the components, aligning the fibers, and of the fiber itself increase quickly as you try to raise the data rate.
Paniccia wouldn’t comment on where Intel stood with respect to Luxtera and other contenders. ”I don’t want to get into the competition thing. This is about lifting all boats,” he says.
Both companies see silicon photonics as an attractive way to use optics to speed up transmission of data in computers. Copper is still the best choice for moving data over a distance of several centimeters, but as data rates increase, the distance over which copper has the advantage is getting shorter, says Paniccia.
”There are billions and billions of bits that you now have to connect and move around,” Paniccia says. ”You can do it in copper, but it’s becoming very painful, very slow.”
For a 46-centimeter copper-based interconnect, 10 to 15 Gb/s is about the practical limit, but silicon photonics can scale to a terabit per second and beyond, he says. There are two ways to increase capacity: One is to add more channels—the combination of lasers, modulators, and detectors. The other is to increase the data rate per channel. A link with 25 channels operating at 40 Gb/s each would provide 1-terabit-per-second transmission. Paniccia says that’s enough to download the entire Library of Congress in about a minute and a half. What’s more, the availability of high-speed and relatively long-distance links could change the architecture of computers. Without the need to place memory close to the central processing unit, for instance, designers could build faster, lower-power systems, he says.
The channels in Intel’s prototype were limited to 12.5 Gb/s, not because of the link itself but because of limitations on the complementary metal-oxide-semiconductor driver circuits that produce the signal in the first place, says Paniccia. He says Intel is already working on increasing how fast those circuits work. Other improvements they’re working on to turn the prototype into a commercial product include improving the performance and reducing the operating threshold of the lasers, developing modulators that operate on less than 1 volt, and optimizing the integrated system for high-volume manufacturing. Paniccia thinks the company will be selling the interconnects by the middle of the decade, with devices operating at perhaps 500 or 800 Gb/s for high-performance computing, 200 to 400 Gb/s for data centers, and 50 Gb/s for consumer electronics.
Luxtera, meanwhile, is in the midst of a five-and-a-half-year, $44.3 million collaboration with Oracle and start-up Kotura to bring optical interconnects even closer to the chip. The companies are working on smaller, lower-power optical interconnects for transmission between chips and between the cores of multicore processors. The program is funded by the U.S. government through the Defense Advanced Research Projects Agency. At the same conference at which Intel made its announcement, researchers from that team discussed their progress in making low-power photonic links for a so-called macrochip, in which a cluster of chips is linked optically.
This article was updated on 24 August 2010.
About the Author
Neil Savage writes about optoelectronics and other technology from Lowell, Mass. In the July 2010 issue he reported on memory chipmaker Hynix’s efforts to adopt Innovative Silicon’s Z-RAM high-density memory technology.