Bye-Bye Bottlenecks

Young inventor speeds chip-to-chip communication 100-fold.

By Al Riske

15.Nov.04--Dr. Robert Drost has been named as one of the world's top young innovators -- individuals under the age of 35 whose work in technology is having a profound impact.

Just in time, too. He's 34.

The honor comes from Technology Review, the prestigious "Magazine of Innovation" published by the Massachusetts Institute of Technology.

One reason: Drost's work in wireless chip-to-chip communication allows data to travel between chips about 100 times faster than the limits of any technology on the market today.

The Wall Street Journal has taken notice as well. Today the paper recognized the project -- the work of Drost and a small team of Sun researchers -- with top honors in its 2004 Technology Innovation Awards competition.

"Instead of accepting classic elements of chip design, Dr. Drost is challenging one of the most basic assumptions of the semiconductor business," the Journal reports.

The new technology is capable of transferring terabytes of data per second. Imagine what it will mean for data-intensive applications such as mapping distant galaxies, simulating protein folding, or projecting the results of medical treatments, including chemotherapy.

In that context, "profound impact" starts to sound like an understatement.

Born and raised in New York City, the son of European immigrants, Drost smiles easily and says he's "proud to represent Sun" on the Technology Review 100, but the distinction doesn't seem to have affected him.

He will be the first to tell you, for instance, that the idea for wireless chip-to-chip communication, known as Proximity Communication, came from a much-admired senior colleague at Sun, Dr. Ivan Sutherland.

Drost's own contributions have been significant, however.

For one, Drost found an ingenious way to overcome the biggest obstacle between the idea and its practical application.

In Proximity Communication, chips are positioned within microns of each other, but not necessarily touching. That permits transmitters on one chip and receivers on the other to exchange data at on-chip speeds without being connected by off-chip wires, soldered connections, or other current techniques that really slow things down.

The first big hurdle: How to make sure transmitters and receivers stay closely aligned -- the slightest shift means they can no longer communicate.

Drost's solution came to him in the usual way: unexpectedly. He had been thinking about how to optimize the size of the receiver in relation to the transmitter, maybe bring in the edges of the plate a bit, make it half as wide. And he was testing to see how far off-center the receiver could be and still pick up the transmitter's signal. He discovered a sort of plateau area where the signal remained clear enough, but it didn't really solve anything.

Then the answer just popped into his head. He wasn't even at work at the time. It was a weekend, and he doesn't remember now what he was doing. He just remembers this sort of mental leap.

I can move it, he realized. Electronically, I can move the position of the transmitter plate.

What?

"I said, If I break the transmitter plates into a series of smaller plates, then I can effectively move the transmitter one block over if I need to. In that way I can extend the plateau indefinitely."

Sure enough, his idea worked. The technique requires something called steering multiplexer circuits to shift the transmitted pattern to match the location of the receiver plates, but it works. Drost and his colleagues have already built and tested the prototype.

All that may be a bit esoteric, but the bottom line is this: Sun holds the all patents for Proximity Communication -- about 16 issued or pending so far, Drost says -- and it's really, really fast.

It's one of the reasons, in fact, that Sun Labs won a $49.7 million contract last year from the U.S. government -- specifically, the Defense Advanced Research Projects Agency, or DARPA -- to help design next-generation supercomputers.

And it's yet another example, Drost says, of Sun's contrarian approach to technology.

The conventional wisdom has been to pack more and more functionality onto a single chip. You get more speed that way, but there's a problem, Drost says. The more you put on a chip, the bigger it becomes. The bigger the chip, the higher the odds of a defect -- a microscopic speck of dust destroying its effectiveness.

"What we're doing that's contrarian here is going in the other direction -- splitting the functionality of the computer into a number of chips and using this proximity technology to interconnect them with really high bandwidth," Drost says.

Smaller chips also mean higher yields in manufacturing. That is, more chips that are free of defects. Which makes them cheaper to produce.

"The big thing, though, is we're getting rid of the wires," Drost says. "By eliminating wires, we get a huge increase in density and the number of communication channels we can put on a chip."

Sun is bucking conventional wisdom in another way, too, and it's in an area Drost is deeply familiar with.

As a doctoral student at Stanford University (while working part time for Sun), Drost was able to engineer some fantastic performance gains by doing what everyone tries to do: push the limits of current technologies.

For his thesis, he developed an interface that could drive data through a single wire on a chip at 8 gigabits per second -- about 30 times faster than in commercial systems of the time.

"That was really pushing the highest speed you could go," he notes. "To get to those speeds, you have to throw tons of transistors at the problem -- a ton of transistors to compensate for the poor quality of the transmission line and calibrate out the distortion you get."

His technique was fast, but Drost could see there was a price to pay. All those transistors consume a lot of power and generate a lot heat. So, if you try to build a large-scale supercomputer that way, he says, you'd have to blow a hurricane through it to keep it cool.

"If the amount of heat you can dissipate is fixed, the only way to increase performance is to use power more efficiently," Drost says.

Which is why he finds Proximity Communication so exciting. No wires means fewer transistors using a lot less power. It also means you can easily add a whole bunch of wireless communication channels to deliver a massive amount of bandwidth.

And it offers real advantages over the approach others have taken to 3-D chip technology, which is to sandwich several chips together.

"It is possible, by stacking chips, to get quite high bandwidth between them, but it doesn't help with the next step," Drost says.

"Say you're to the point where you can laminate a few chips together. That's great, but if you don't have Proximity Communication, what do you do now? How do you get this collection of chips to communicate with another collection of chips at high speed?"

With Proximity Communication, you can do both.

"We think it's valuable to laminate a couple of chips together, where it makes sense," Drost says. "You can actually increase the density of data and processing in the system. But then the outside chip will have Proximity Communication technology on its outside face and that's how it will communicate with the rest of the system."

What Drost and other members of the team are trying to do is build a supercomputer with hundreds of thousands of processors talking to each other with bandwidth that's comparable to on-chip cache.

"It's a whole different computer when you can do that," Drost says.