Making Wires Cool

"Cool Wires" is a Technology Ahead of Its Time, But Not for Long

By Al Riske

19.July.07- As microprocessors have gotten smaller and faster, the wires within them have gotten smaller and slower.

"A wire is kind of like a garden hose. A narrow hose will give you less flow than a big hose," says Ron Ho.

"Today we manufacture chips using a 65 nanometer process. We're going to move eventually to a 45 nanometer process and to a 32 nanometer process. As wires get smaller, my garden hose is getting narrower."

Ho, a Distinguished Engineer in Sun Labs, describes himself as "kind of a wires guy." His job is to look ahead, and here's what he sees: Slower wires and a lot more of them.

There are a great many wires linking the eight cores on today's UltraSPARC T1 processors, for example.

"What happens when we go to 64 cores?" Ho asks. "What if we go to 128 cores?"

Of course, the wires connecting those cores all carry energy back and forth.

A small amount of power today multiplied by a large number of such wires in the future becomes a big problem," Ho says.

Which is where his "Cool Wires" project comes into play.

Ho joined Sun to work with Ivan Sutherland and Robert Drost on a new technology known as Proximity Communication -- transferring data between chips without wires at incredible speeds. Terabytes of data per second, in fact.

"Robert and I and Tarik Ono [a Sun Labs colleague] were looking at this once and had the brainstorm: 'Hey, we can use this not only between chips but also on a chip,'" Ho recalls.

Simply put, the key concept behind Proximity Communication is a natural phenomenon called capacitance: "If you place two plates of metal next to each other, a voltage on this plate will cause the other plate to have a similar, though reduced, voltage. That's data transmission," Ho explains.

As a result, chips placed in close proximity can communicate without wires.

Within a multicore chip, you still need wires to connect the cores, but Ho has shown that capacitors can have a beneficial effect there as well, making all those wires 30 percent faster -- while using one-tenth the energy.

Hence the name Cool Wires.

So how does it work? Ho suggests we take a step back and consider a couple of simple analogies. Like the difference between shouting and whispering. Or the difference between dried pasta and a wet noodle.

"Let's say this is my data pattern: 1010001011101. Something like that. If I send that over a long wire to the receiver, the entire wire moves up and down, metaphorically speaking. Each time the wire moves up, it uses energy, energy that has to come from the wall socket or a battery. So the more you use the wire, the more times it moves up and down, the more energy it consumes," he says.

"Now if I put a capacitor here, it reduces the amplitude of the communication. If one side moves up, the other side moves, too, but not as much. Zero to one volt on this side, zero to 100 millivolts on the other -- one tenth as much. Now that's interesting. By just adding a capacitor, I can save 10x the energy."

So far, so good. But what started as a shout is now a whisper and, Ho concedes, that makes a lot of people nervous.

"You can sit there and hear me whisper, but then let's go into a crowded room. Can you still hear me? Microprocessors are noisy chips, and so a large portion of the work is making sure we can do this in a way that guarantees the receiver can hear the correct data," he says.

"What we're doing is putting a capacitor on one end of the wire, making this very low energy, at the cost of having to magnify it on the other end. That's not trivial but these magnifiers exist in our chips all over the place. They're called sense amplifiers. So we use the exact same circuits we have in our caches, our register files ... they're tricky but we have people who know how to design these circuits."

Ho notes that people have been talking about doing this -- low-swing signalling -- for years. It was, in fact, one of the techniques he outlined in his Ph.D. thesis, "The Future of Wires," written soon after he understood how a few wires could limit the speed of an entire microprocessor.

"In commercial systems, nobody has really done low-swing signalling," he says. "Test chips have been built at universities, but in the real world of building products it's hard to get anyone to spend the money."

That's because, up till now, low-swing signalling typically required an additional, custom power supply.

"All the stuff you need to deliver power to a chip is so expensive and so complicated that it would double the cost of power delivery, double this, double that," Ho says. "That's why the capacitor is so important. Maybe it isn't as efficient as using a second power supply, but it gives me the low swing without having to build that second supply. I get this low swing for free, as it were."

That's half the neatness of the Cool Wires solution.

"The other half is that this capacitor actually speeds everything up," he says. "It's something I didn't actually notice at first."

Here he applies the noodle analogy.

"Pretend this wire is a piece of dried pasta. If I want to send a one, I have to lift up one end of the piece of pasta. A zero, I bring it back down. Up, down. Up, down. It's rigid so the whole thing is going to move together, and the far end of the pasta moves at the same time as the near end. That's how wires used to be. Very fast," he says.

"Today wires have gotten smaller and slower so they're more like cooked pasta. Now when I lift up one end, this wave will eventually make its way down to the other end. But it's a limp noodle. It's going to take some time. The whole thing will move slowly. But by using a capacitor, we can accentuate quick changes in voltage. It's like whipping your wrist and giving a good snap to the wet noodle. The pulses - high, low, high, low -- travel a lot faster."

Another important consequence: Making long wires faster ends up saving even more energy.

"A long wire may be so slow I wouldn't be able to use it. It would limit the speed of my microprocessor. So the standard solution today is I break the wire in half and put another driver (a repeater) in the middle to give it a kick in the pants. That's fine but I've got to allocate space for the repeater, I've got to burn the power of the repeater, and make more complex design decisions," he says.

"But if I just use a capacitor on the long wire, maybe now it's fast enough that I don't need that repeater. So it's a simpler design, I burn less power, and I don't need to allocate space for the repeater."

One last point about the design, inspired by the wires themselves.

"A wire in cross-section looks like a rectangle -- taller than it is wide. Why? As we scale technology we want to make things more compact, which means the wires get narrower and narrower. (Think garden hose versus fire hose. A lot less flow.) In response, fabrication people said, 'I can help you a little bit by making them taller.' It's only a partial help but it's a help, so we do it. Now wires are two or three times taller than they are wide," Ho explains.

"All the wires across a chip will have this arrangement -- they are very tall and not very wide and they get exposed to other wires along this very tall side. That causes noise as the information each wire carries inadvertently leaks over to its neighboring wires."

Ho jokes that he can imagine the wires saying, "Shut up! I'm trying to do my own data communication. I don't want to hear your stuff."

He says there are circuit techniques and layout techniques for dealing with the problem -- also outlined in his doctoral thesis -- but he looked at the situation again and saw an opportunity.

"I realized I could use this exposure area to be my capacitor," he says.

In other words, the intrinsic capacitance of the copper wires could be put to good use. No special processing required. No need to use transistors in some clever way.

"The wires can be the capacitors, automatically, just by laying them out in a specific pattern," Ho says, drawing what looks like two face-to-face pitchforks on the whiteboard in his office. The two forks are jammed together, with their tines interleaved, and their handles, representing the rest of the wire's two sides, pointed away into space.

"Granted, this structure is not small and that's one of the things we're working on, making this as compact as possible. But it's wire. If I had to do it out of transistor, I'd have to allocate space for the transistor and that's not a fun thing to do either."

This is not a technology that is really needed today, Ho says, but he wants to make sure it's perfected in time to meet a crucial coming need.

"As the folks in the processor group start pushing for more and more cores, we hope this is a technology they can put into some of their plans," he says.

"I'll be the first to point out that they can't just take it wholesale and use it without thinking about it. They do in fact have to worry about this amplifier circuit. They do in fact have to worry about noise. But we've built chips and showed that it works, and we want to work with them to re-prove it. We're talking with them now about putting this exact circuit on some of their test chips."