Sun Bags $44.3 Million DARPA Contract for Funky Chip Interconnect
Published: March 27, 2008
by Timothy Prickett Morgan
While the company is perhaps best known for its software innovations, such as Network File System and the Java language and related runtime environment, Sun Microsystems and the many companies it has acquired do a lot of research in hardware, particularly in chip architecture. The engineers have been working on a new chip interconnection technology called proximity communication for the past couple of years, and this week, Uncle Sam ponied up $44.3 million in cash to help Sun continue its research and perhaps, just perhaps, radically change the way computer systems are manufactured.
As 2007 was coming to an end, I talked to Hans Eberle, a distinguish engineer at Sun who is working on proximity communications and who was at pains to make sure people understood that this was still primary research and not part of any impending Sun systems launch. The new technology that Sun has been cooking up in the labs seeks to link chips together into multichip modules (and therefore systems) using a contactless interconnect. The contract that Sun has just inked with the Defense Advanced Research Projects Agency, the arm of the U.S. military that funds computer and networking technology research (such as the products that today make up the Internet), calls for Sun to work the kinks out of proximity communications and to also work on a technology called silicon photonics, which mixes electronics and lasers to create communication links between components. The end result, DARPA hopes, is to be able to create on-chip optical networks, which have a lot more bandwidth than wires can handle and do so at a lot lower energy levels, too.
The idea is not that complex, although the technology certainly is. Right now, every chip manufacturer in the world is riding the Moore's Law curve, trying to get as many components onto a single piece of silicon as they can because this is the only way to keep boosting performance at the system level. But as transistors shrink and devices have billions of transistors, getting yields on such physically large devices is a bit of a problem. Yields would be higher on smaller chips with fewer components--the yield of total cores on a platter of eight-core chips is a lot lower than it would be for processor made up of individual cores, for instance, because a certain number of boogers on the platter would tend to knock out eight times as many total devices. The problem is that connecting chips together means putting things in sockets or expensive multichip module packages that are expensive and cranky.
Proximity communication takes the area ball bonding used on the bottom of chips and not only shrinks it, but gives it a Lego-like interconnect structure. Instead of putting the ball grid on the bottom of the chips, Sun is putting links on the edges, and chips can be stacked in arrays to link to each other without contacts using capacitive coupling between the proximity interconnect pads. The transmitters and receivers in Sun's research--called micropads--can operate at a range of between zero and 20 microns, and the chips themselves are only 300 microns thick. The scale of the proximity interconnection is small--about 15 microns compared to the 120 micron spacing in ball bonding schemes, and Sun thinks it can put chip I/O up to 10 terabits per second per square millimeter, compared to the 10 gigabits per second per square millimeter of overlap in area ball grid approach used for chip-to-chip links today. This is a huge improvement.
But proximity communication is more than just increasing bandwidth. Because the links are more like wireless links than wires, chip packages that have multiple elements--multiple cores, memory controllers, networking NICs, and so forth--can be taken apart and fixed. And to link them all back together, you just drop them into the package (which has to be cleverly designed, mind you) and shake. The chips fall into their proper places in the package, everything links up, and you plug it in and turn it on. Kinda cool, right?
What DARPA wants Sun to do as part of its five-and-a-half year contract is to take proximity communication technologies and merge them with the work that DARPA has been funding as part of its Ultraperformance Nanophotonic Intrachip Communication program, which has been exploring another chip interconnection technology called silicon photonics. That is just a fancy way of saying really tiny laser. And where Sun is using capacitive coupling on the proximity communication scheme in its research, DARPA is asking Sun's Microelectronics group and its Sun Labs research arm to come up with a way of putting laser communications between the bits. Laser light has much lower latency and much higher bandwidth than electrons moving in any wire, large or small. And, because there is no soldering involved, systems should be able to snap together in a contactless way. The scale of a system--which Sun has started calling a macrochip because, well, Sun has always been like this--could be immense.
While such research is interesting, and Sun is very keen on flexing its brains in the high performance computing space, I do recall in the mid-1990s being told that the UltraSparc-III machines Sun would deliver by the late 1990s would scale to over 1,000 processors and put anything else anyone can conceive of to shame. Sun is being a lot more realistic this time around, which is again commendable. Sun's Ron Ho, a researcher at Sun Labs who is a leader on the DARPA project, said that there is a 50 percent chance this project fails. Honesty is always the best policy, and it is also always a good idea to get the government to pay for the research. This strategy sure has worked for all the other supercomputer makers in the world.
TSMC Chosen as Future Processor Foundry by Sun
Sun Delays "Rock" Sparc Machines Until 2H 2009
Rock and Tukwila Are the Stars of ISSCC This Week