This week, the Defense Advanced Research Projects Agency (DARPA) awarded Sun Microsystems a five and a half year $44.3 million dollar contract to research and develop commercially viable optical interconnects for next generation multiprocessor systems. According to an article by John Markoff of the New York Times, the government selected Sun’s proposal over three competing bids from Intel and HP; IBM; and MIT. For the contract, Sun will be partnering with Luxtera and Kotura, two silicon photonic vendors, to leverage their optical expertise. In addition, Stanford University and the University of California, San Diego will be collaborating on the work.
The DARPA program — called Ultraperformance Nanophotonic Intrachip Communications (UNIC) — has the goal of “demonstrating low power, high bandwidth, low latency intrachip photonic communication networks designed to enable chip multiprocessors with hundreds or thousands of compute cores to realize extremely high computational efficiency.” The idea is to use the high bandwidth and low latency characteristics of photonics to create ultrascale chip grids — ‘macrochips’ as Sun calls them. Combining scads of processors into one big virtual processor via high bandwidth links circumvents the yield problems encountered when chipmakers try to make the devices too big.
Part of the UNIC work will be derived from research connected to Sun’s High Productivity Computing Systems (HPCS) Phase I and II work. When DARPA eliminated Sun from Phase III of the High Productivity Computing Systems (HPCS) program in November 2006, it seemed as though some of the company’s research developed under the first two phases of the program might go for naught. But it looks like at least one piece of the HPCS work will provide a jumping off point for this new program.
Sun’s UNIC model is based on “proximity communications” technology that the company was pushing in their HPCS submission. That technology, which Sun says it is continuing to work on separately, allows adjacent chips to talk to each other via electromagnetic fields, instead of unwieldy, low bandwidth physical connections. One disadvantage of the proximity communications model is that it still relied on copper wires on the chips themselves to shuttle data around. So even though the chip-to-chip proximity communication latencies are low, for a really large grid of chips (think 100 x 100 processors), the on-chip latencies add up as you traverse the larger macrochip.
That’s where silicon photonics has the advantage. The new project will use the proximity model, but substitute light waves for the longer wavelength electromagnetic waves to talk across the intrachip gap. And because optical waveguides can be built into the chips themselves, all communication across the grid can move at the speed of light — at least ten times faster than electric signals over wires. Essentially, they’re proposing a device in which they can route data across the entire chip grid via a series of photonic highways and bridges. That solves the latency problem. According to Ron Ho, one of the three Sun Distinguished Engineers on the project, they’ve already demonstrated this in the initial phase of the program.
With recent advances in nanophotonics, chipmakers can now pack multiple wavelengths of light into a single silicon waveguide, which gives you lots of bandwidth. On a per channel basis, a one quarter micron waveguide can handle up to 16 wavelengths of light, each achieving 15 Gb/s. So a single waveguide could handle over 200 Gb/s. If you aggregate 1,000 channels, running along the edge of the chip you can achieve 200 Tb/s.
The biggest challenge, says Ho, is energy consumption. To make any of this worthwhile, low-power photonics are required — something currently missing from today’s solutions. Intel’s approach, which integrates the lasers themselves on-chip, is rather power hungry, so Sun is using the more conventional discrete laser approach. But even the power just to drive the on-chip optical links is rather high — even slightly greater than that for serial communications used for memory communication channels, which is in the 20-25 mW/Gb/s range. For the UNIC work, Sun is aiming to lower power consumption by two orders of magnitude (100x lower).
According to Ho, the program was set up with very aggressive milestones. Even if they miss some of those, the Sun researchers believe they’ll still be able to derive commercial success from the work. If the company manages to incorporate some of this technology into their UltraSPARC line in the next few years, they have a real shot at changing the game for high-end computers. But Ho admits they probably only have a 50-50 chance to achieve all the goals specified in the program.
“If you’re going to be doing a project that’s funded by DARPA, it had better be something sort of crazily aggressive,” says Ho. “If it’s something that’s conservative and you’re almost completely sure you’re going to succeed, well that’s probably not something DARPA should be funding.”