DARPA taps Sun to take microprocessors to 'macrochip' level

When it comes to computing technology, the research goal of the Defense Advanced Research Projects Agency can be summed up in two words: power and speed.

In the latter category, Sun Microsystems announced that DARPA has awarded it up to US$44.3 million in funding to do research on the use of optical technology to speed up communications between different microprocessors in a system.

The project, which is scheduled to take five and a half years to complete, is aimed at developing what Sun described as virtual macrochips: arrays of low-cost processors that function as one device and can deliver increased computing performance with the help of silicon-based optics.

Chips that are soldered together wire to wire typically communicate at between one-tenth and one-twentieth the speed of light, whereas optics can support connectivity at a near light speed. The idea of using optical communications to increase the speed at which data moves between processors isn't new. Just last week, in fact, IBM announced that its scientists have built a switch that uses pulses of light to control the flow of information on chips, a development that it touted as a step forward in efforts to create on-chip optical networks.

But Ron Ho, a Distinguished Engineer at Sun who is part of the team assigned to the macrochip research project, said that a key issue for the researchers will be reducing the amount of energy consumed by the use of optics in chips. "You can't exploit the power of optics without bringing the power way down," Ho said. "That's the risk that DARPA is trying to address with this program."

As part of the project schedule, the Sun researchers have set goals to continually reduce the amount of power used in what is, essentially, a form of wireless communications between chips. A related goal set by both Sun and DARPA is sharply reducing the cost of the optical connections.

Aside from making it easier to create supercomputers, the macrochip concept should help to lower technology costs, according to Ho. For instance, he said that if there was a problem with any processor in a multi-chip array, it could easily be removed and replaced. That isn't the case now with chips that are physically connected to one another.

Macrochips using optics also could bring enormous performance gains to supercomputers, which is another goal of DARPA. The agency has a separate project under way to boost the size of supercomputers, and in 2006 it awarded IBM and Cray US$244 million and US$250 million, respectively, to develop petascale systems. The two companies are scheduled to have prototypes ready by 2010.

Sun's work on optics likely won't result in the addition of new technologies to the company's business systems for several years, at least. "It's a very interesting area of research, and I think two or three years would be extremely aggressive for this to turn up in a product," said Nathan Brookwood, an analyst at Insight 64.

Building multicore chips is also increasing the size of chips. But Brookwood said that optical technologies could enable chip makers to separate memory from the processor itself on a chip and rely on the interconnect to boost performance instead.