SAN JOSE, Calif. The computer industry wants optical interconnects in its future multicore processors, but just how it will get them remains unclear. That was the conclusion of a panel of experts from Hewlett-Packard, IBM, Intel, MIT and Sun at Photonics West here.
Several factors are driving the need for silicon photonics in computing. However, experts agreed there is no clear light source to drive on-chip optics, and the technology needs to get to significantly lower heat, power and cost to be viable.
"We need a byte [of bandwidth] per Flop [of performance], said Ray Beausoleil, an optics researcher at HP Labs. In a keynote here he said the computer industry sees on-chip optics as a way to maintain a relatively flat programming model for tomorrow's massively multicore processors.
"We found we had to maintain that byte-per-flop metric," agreed Jeffrey Kash who manages an optical research group at IBM, referring to the company's work on a massively parallel supercomputer it is developing as part of a government research program.
"The perception is the need for bandwidth is the reason for making the shift to optics, but it's also the freedom to repartition the computer architecture to get access to more memory and new abilities to cool the system," said Mario Paniccia, director of Intel's silicon photonics lab. "It's not just the next speed bump," he added.
The panelists agreed that no clear silicon light sources are on the horizon
"Teams are working on silicon lasers in many companies, but even if the technology was working today it would still be five years from commercialization," said Paniccia. "Many of us our hopeful, but [it's hard to find] the milliwatt efficiencies we are looking for with a decent conversion rate on infrared wavelengths for transparency," he added.
Kash of IBM noted Intel has taken an interesting hybrid approach using III-V materials for the gain function in ways that could be packaged to be compatible with silicon photonics.
"Silicon lasers are a ridiculous concept if you want to build something real," said Eugene Fitzgerald, a professor of materials engineering at Massachusetts Institute of Technology. He said a practical solution "will be III-V devices on-board, and then you have to find an inexpensive platform to get the heat out."
"We strongly prefer not to have the light source on the chip—or even in the data center-- because we don't want to cool it," said Ashok Krishnamoorthy, a distinguished engineer in optics at Sun Microsystems. "I'd rather have light pumped in from outside if I could."
"I'd prefer to get [the light source] away from other heat sources so we could cool it separately," agreed Kash. IBM researchers have found on-chip optics "need temperature stability within a degree which is not practical, so we need to make these systems more temperature tolerant," he added.