Luxtera Lights Up Sun
The company is working on 40-Gbit/s interfaces -- four lanes of 10-Gbit/s apiece -- for Sun and will demonstrate some of its technology at this week's SC|05 supercomputing show in Seattle.
You might recall Luxtera as the startup claiming to outdo (Nasdaq: INTC) and others in the silicon photonics game. While Intel is pursuing a silicon laser, Luxtera is skipping that piece, instead targeting the other optical parts such as modulators and transimpedance amplifiers. (See Luxtera Chases Silicon Photonics.)
The goal in all this is to build optics using complementary metal-oxide semiconductor (CMOS) methods -- that is, to make them from the same stuff as commonplace chips, rather than from materials such as indium phosphide (InP). Such a step could lower the cost of photonics dramatically, as well as increase manufacturing yields.
Moreover, CMOS carries the promise of integration, further shrinking the size and power requirements of optics. Systems vendor Infinera Inc. has shown some extensive photonic integration in InP. But to do it in CMOS could open a greater spectrum of commercial possibilities. (See Infinera Declares WDM War.)
Alas, Luxtera doesn't want to apply that dream to telecom circles. It's looking like the company's first commercial applications will be for interconnects inside a computer, because that's where the money appears to be. Sheesh.
"There are two primary areas of application for the technology right now. One is DWDM for datacom applications. The other is for vanilla 10-Gbit/s data communications," says Alex Dickinson, Luxtera CEO. Luxtera's first commercial chip, slated to sample in February, will tackle the latter case, he says.
Luxtera is keeping its telecom options open by concentrating on wavelengths in the 1550nm range. For today, though, the spotlight is on the company's chances in supercomputing.
Luxtera's announcement today is tied to the U.S. government's High Productivity Computing Systems (HPCS) program, where Sun is competing with Cray Inc. (Nasdaq: CRAY) and IBM Corp. (NYSE: IBM) for the final contract. The nine-year Defense Advanced Research Projects Agency (DARPA) program started in 2002, with the final vendor or vendors chosen in 2007.
The computers involved here are ridiculously large. Not surprisingly, something that big tends to be inefficient. "It's not uncommon for an application to only realize 5 percent of the peak performance of the system," says Mike Vildibill, Sun's director of HPCS product planning.
Part of the problem is the increasing number of processors needed inside large computing systems. More processors means more interconnects, or even more wires connecting racks or linking far-apart chips. It all adds up to more delays.
So, a key to Sun's HPCS architecture is the interconnects. Sun's HPCS machine is partitioned into modules of densely packed chips. Modules close to one another can speak via a Sun-designed, high-speed electrical interconnect. But for Hail Mary kinds of distances, Sun wanted to shift to DWDM optics, starting at four wavelengths of 10-Gbit/s apiece with the potential to use dozens of wavelengths later.
What made Luxtera right for the job was its CMOS technologies, Sun officials say. That gives Sun the potential to combine Sun's special electrical interconnects and the longer-range DWDM links onto one chip, further shrinking the size -- and the power requirements -- of those chip modules. Eventually, the result could be "tens of terabits per second" worth of interconnect per square millimeter, Vildibill says.
Luxtera is far from unique in trying to apply silicon optics to computing. Just this month, IBM announced lab results that slowed light significantly in silicon, which could become a step towards processing information optically. (For other examples of research in this area, see Scientists 'Stop' Light In a Solid and Storing Light).
"People are working on various building blocks" related to optics and computing, says Fred Zieber, an analyst with Pathfinder Research. "IBM said one of the reasons they were working on it was as a potential chip-to-chip interconnect within a supercomputer. You're still a few years away from functional applications."
— Craig Matsumoto, Senior Editor, Light Reading