I just got off the phone with Ron Ho, a distinguished engineer at Sun research. Here's some further details about their $44-million darpa optical interconnects project.
First, some shocking frankness: Ho gives Sun about a 50-50 shot at meeting Darpa's goals. The upside to that, is that the goals are so outrageous, that Sun would probably be very, very happy to commercialize the failure.
Here's the goal, by the way: today's copper-based interconnect technology, say Rambus or DDR, costs you 10-20 mw per gigabit per second. For otpical systems, it's little higher than that, says Ho. "We intend to get that down by a couple orders of magnitude," he says. Bandwidth, would then be free, and system architects and software guys would go berserk.
My first question was why do you think Sun's team won? Ho qualified his answer with: "Darpa is pretty opaque; you don't often know what pushes their buttons." That said, he was convinced that Sun's team won because it brought a complete top to bottom approach, from the individual components to the large-scale systems. Ho is a circuits guy and the other two leads, Ashok Krishnamurthy [apologies for any spelling errors] and John Cunningham [same apologies], are fundamental optics guys. So together they brought the device-level view, but Ho says they brought in their systems architects, too. They were able to look at the question at the level of "How would we modify Solaris [Sun's servers] to take advantage of this?"
OK, OK, HP/Intel and IBM are also teams that could bring the device to systems view, too. So maybe it was Sun's proximity communications technology? Well, maybe. "We know that Darpa likes proximity communications." They did fund it in the first place.
Ho says that in thinking about the question, they started with the proximity technology linking two chips and then asked how it would work if you had a 10 x 10 grid of chips. "Sending a bit of data from one corner to other the latency might add up to a couple 100 clock cycles." Well, that's no good.
So the idea is to make a sort of optical proximity connection. The chips would shine and detect light from their edges without the need to string an optical fiber or other waveguide between them.
So what's the plan, Sun? Well, there's 4 or 5 different plans, according to Ho. They and all their partners will be trying lots of versions of this basic idea and the technologies that would enable it and choosing the best combination they can find. By the way, the team members are: Sun, Kotura, Luxter, Stanford University, and UC San Diego. I left out the last two because the New York Times did (and Sun's press release left out everybody but Sun).
The universities, says Ho, are important for providing the "whacky and crazy ideas."
I'm going to talk to Luxtera tomorrow.