'Tiny Bubbles' Could Empower IBM's Next-gen Power6 CPU

While IBM today unveiled a new dual-core Power6 processor for its new System p570 server, resulting in what the Transaction Performance Processing Council confirmed to be the best performance score per core recorded thus far, a manufacturing advance still in the works there could give its next generation Power6 even higher performance on account of improved signaling.

The advance was poorly explained by the general press a few weeks ago: In actuality, it's a manufacturing process that enables vacuum gaps to form in specifically designated segments of chips, reducing or replacing the need for microscopic glass insulators around copper wires. The tiny bubbles, or "airgaps" as IBM calls them (even though there's no air in a vacuum), serve as a far better insulator. In so doing, they improve the signaling ability of chips - by their very presence, enabling them to be sped up.

The new Power6 employed in the latest System p570 is clocked at a staggering 4.7 GHz. But by 2009, the company's Power6 chips could be clocked still faster, with airgaps helping not only insulate the wires better but reduce leakage - the cause of both signal noise and heat.

As IBM's brochures appeared to explain at the time - and as news services dutifully repeated - the vacuum gaps actually sped up the flow of electricity. We didn't realize that was possible with anything smaller than, say, FermiLab. But as the chief scientist on the airgap project, IBM Fellow Dan Edelstein, told BetaNews, the airgap process doesn't speed up electricity. Rather, it expedites the signaling capability of an electric circuit.

IBM Fellow and senior scientist Dan Edelstein. (Courtesy IBM)

"What we're talking about are the multiple layers of copper wires above the silicon transistors on a chip," Edelstein told us. "Microprocessors have a layer of transistors embedded in the semiconductor, because you need a semiconductor to make a switch. And then you've got to connect them all together, and it turns out that that takes many, many layers of wiring to complete all the circuits. So the copper wires have to conduct signals, and those signals involve charging the wire from a 0 to a 1, or from ground to a certain voltage, or vice versa - discharging it from a 1 to a 0."

If you've never thought of it this way before, consider the signal carried by a microprocessor as a square wave...though not perfectly square. It takes time and effort to move the wave from 0 to 1 and back again, and that expenditure is typically measurable.

"To get charge on and off of that line, you have to send current into it," Edelstein continued. "And if it's a long line, you also have to propagate that pulse a long distance - you have a so-called transmission line effect. So it's not the speed of the electrons, it's the growth or charging of the entire signal from a 0 to a 1, and that's limited by the resistance of the metal and the capacitance of the wire. And less is better."

A decade ago, Edelstein and his IBM team led the charge to replace aluminum used in transistor wiring with copper. The reason is because aluminum could only shrink so far before its lack of durability made it unusable. At the same time, resistance has to be reduced in order for size to be reduced as well.

"So now we've been banging on the insulator between the wires, to try to reduce the capacitive coupling from one wire to its neighbors," he explained. "What that capacitive coupling does is sap energy from that signal on the main line to its neighbor, so you have to use more energy, or more charge in the first place just to get the desired result, and it also does affect the propagation of that signal, which is a pulse, down along the line. Higher capacitance means a slower rise of that signal at the far end."

Here's a more practical view of the problem: I've tied a kind of "copper wire" between a fairly reliable sender of signals and a reasonably sensitive receiver: a young maple tree. For me to shake the tree using this wire, I need to pull it down fairly low to "0," raise it fairly high to "1," and snap it down sharply to "0" again.

With high resistance, I have to snap from above my head to below my knees just for the wave I produce to make an appreciable bounce off the tree. So while optimally my signal should propagate quite nicely in a pipeline of uniform height between 0 and 1, from me to the tree, in actuality, my signaling area is more of a narrowing tunnel.

Now, if I were in the company of multitudes of other trees, and an equal number of signaling colleagues, capacitance might become a problem. "If you have too much capacitance, the size of that wave decays and gets smaller, so by the time it hits the tree it's too small to switch the tree from 0 to 1," stated Edelstein. "Also, the speed at which that wave gets from your hand out to the tree is slowed down. In this case, the force, the energy of that pulse is siphoned parasitically to neighboring wires that you don't want to get any of your energy. We call it parasitic capacitance."

By reducing the metal's propensity to store electric charge, more of that charge gets distributed down the wire rather than leaked to other wires parasitically. The perfect insulator is a vacuum, but alas, there's no way to embed wires in a perfect vacuum. However, billions of tiny nano-vacuums in the vicinity of the wire can help.

The measure of a substance's relative capacitance is its dielectric constant. This is the number IBM scientists want to drive down. A perfect vacuum has a dielectric constant of 1. By contrast, water - which is quite terrible - would have a dielectric constant of 80; and worse still, surrounding a metal with itself would yield an infinite value for a dielectric constant.

"For 40 years, we've been using silicate glass as the insulator that all the wires are embedded in," IBM Fellow Dan Edelstein told BetaNews. "And that has a dielectric constant of 4...In the last few years, with great effort, the industry has pushed that number down to 3 by doping the glass with carbon; and as you push it down, you also make the glass much more fragile and brittle, to the point where it's questionable how much farther we can go. The airgaps immediately, in one step, use the same strong dielectric by cutting slots into it, reducing it down to 2. And the dielectric constant in the gap is actually 1, which is the lowest anybody can go in nature - it's the absence of anything, a vacuum. But since there's still insulator above and below the wires, the effective or overall dielectric constant adds up to about 2."

Next: Puncturing Holes in Your CPU Without Weakening It