There's no fundamental barrier to the growth of the integrated circuit, according to Carver Mead, a seminal thinker whose work has spanned much of the growth of electronics. In a wide-ranging interview with EE Times, Mead shared his thoughts about how advancements in arenas as far afield as biology will help push semiconductor technology beyond the widely perceived limit of the single-electron gate.

Mead's optimism has proved well-placed in the past. In 1969, he projected that improved processing techniques would permit semiconductor features as small as 0.15 micron. He went on to predict that future ICs would house millions of transistors packed together via very large-scale integration (VLSI). In the book Introduction to VLSI Systems, Mead and co-author Lynn Conway described design techniques for just such complex circuits.

Today, Mead believes designers can leverage lessons from biology to bring the IC to the next level of maturity. Now a professor at the California Institute of Technology (Pasadena), Mead became fascinated with issues in analog VLSI designs based on his observations of the workings of the human nervous system. Since then, he has created both a silicon retina and a silicon cochlea (inner ear). A crowning achievement was the 1989 publication of Mead's Analog VLSI and Neural Systems.

EE Times: What are your views on the "atomic wall" down below 0.05 micron? Will scaling theory finally fail-is the atomic wall ultimately insurmountable?

Mead: There are no [insurmountable] brick walls; there is always a way around them. Fifty years ago, no one had a clue that scaling theory would result in picosecond timing and the incredible densities of today. Even if our densities were frozen today, we would still have another 50 years of innovation ahead of us.

If we continue to lean on just the technology we know today-make better fab equipment, make the oxides thinner and the features smaller-we'll get down to 0.03 micron by just following our noses. We have a wonderful depth in and around the way we do things today, and we can't overestimate how important that is. That depth of knowledge will carry us much further than one might think.

EE Times: What might prolong the life of the IC?

Mead: The use of insulated substrates will easily take us below 0.03 micron. The next step beyond that will reduce the tunneling current through the gate oxide. Tunneling makes transistors into a kind of triode, which is the ultimate limiting factor of current processing technology, along with punchthrough in the channel.

There is no question that the physics allows you to go beyond what our current picture contemplates. For instance, we might not build chips in layers like we do today-who knows? The future will grow from the stuff that sounds craziest today.

I have a new electromagnetic theory that starts from basics in quantum mechanics that have only recently become well-understood. Engineers tell me it sounds crazy-but that's what they told me in 1971, when I predicted 0.15 micron.

EE Times: Isn't the single-electron gate the limit?

Mead: Well, the quantization of charge is certainly an important aspect of physics that needs to be taken into account, but that doesn't mean it's some kind of barrier. I consider it more like an opportunity: The natural quantization of charge is a good thing because you would like it to be in nice discrete levels.

In particular, if you have a mechanism for discreteness that is given to you by physics, then that is much preferable than building discreteness into some darn flip-flop. Why not use the physics of the materials to give you 1s and 0s? That would make more sense, wouldn't it?

So if I were going to make a digital device, I would view the natural discreteness of a physical property as an opportunity, not a problem. That kind of optimistic thinking will lead to a whole new class of devices that will no doubt be mostly fluff, but a few good things will come out of it.

The way electrons work is that if you confine them to a smaller space, the spacing between their energy levels gets larger. So using the quantization of the materials, instead of building a flip-flop, as the natural way to make a binary system actually gets better as things get smaller. When you get down around an angstrom, the spacing gets to be on the order of a few electron-volts, and now you're talking.

So an angstrom is kind of a magic number, because things get a lot better there. At an angstrom, you have a nice, convenient energy space that thermal energy can't corrupt. But that is a different paradigm than we are using today.

It's not unknown. The physics is known, and there are no surprises; it's just that we don't normally think about that scale as being better. We think of it as a problem. But we will be much better off thinking of it as a challenging opportunity.

EE Times: As we get down to these unbelievably small integration levels, will we need a new breed of design tools to deal with these "opportunities"?

Mead: As we get down to smaller feature sizes, we have tried to cram more and more stuff onto chips, making interconnectivity more of an issue than feature size. What we have done is take the old, lock-step synchronous paradigm and just flogged the heck out of it. We've made bigger wires to run those clocks around, and we run the clock faster and faster, and we pump more and more power into the clock lines. We've just beat the heck out of this poor little thing.

Now we've got six layers of metal; soon we will have 12 layers of metal. That's all fine, and it all helps, but the capability at the elementary circuit level is vastly more powerful than we are taking advantage of today.

So to return to your question, there are incredible new advancements to be made if we just paid some attention to design methodology. We have taken this methodology for making synchronous circuits, which was perfectly appropriate when the time to communicate across a system was short compared with the time it took the elementary circuits to work, and we extended that methodology to the time where now it is just the reverse-now the time to communicate across a system is long compared with the time each individual device needs to work.

So now we are waiting around for communication to come across wires. That suggests some architectural thoughts as to how can you arrange computation so that it is more localized.

EE Times: So we really have to think of the problem in a whole different way.

Mead: The limiting factor here is our own imagination, not the nature of computation. That's how I got into this neurocomputing stuff, because the brain is a computing structure that is much more distributed than anything we build today. I found it absolutely fascinating, and I've spent 15 years of my life building that stuff and trying to figure out how it worked.

My belief now is that there is much to be learned from the organization of neurocomputing that is applicable to any kind of computing, including both analog and digital: namely, that we should be distributing computations in such a way that local operations can be self-timed.

Also, we learned from neurocomputing that we should not require that all the precursors to a computation be completed before the computation is allowed to proceed. The idea of partial data and all that kind of stuff could play in any kind of information-processing environment; we just need to think of it that way. The nervous system always operates on partial information; its basic assumption is that you start with no information, and anything beyond that is something, and it is a heck of a lot better to have something rather than nothing.

But what the digital paradigm is based on is, "I've got to have all my arguments before I start any operation"-the opposite end of the spectrum! So maybe we can think of some way to have some middle ground there.

EE Times: Then you would have an approximate answer?

Mead: Yes. Well, there is nothing wrong with that.

EE Times: The paradigm you allude to-computing approximate answers from incomplete information-sounds like fuzzy logic.

Mead: Yes, fuzzy logic is a very brain-dead version of what I'm alluding to.

By that, I don't mean there was anything inferior about fuzzy logic. It was a good first step toward doing something of this kind. It allows more or less continuous values and a variable number of arguments. It's a way of building a fuzzy analog thing with digital stuff, and that's fine, but it is only a baby step toward what could be done if we were really smart about it. And there are smart people out there, and they will figure out this stuff.

EE Times: So are higher densities irrelevant? Is a paradigm shift what we really need?

Mead: Don't get me wrong; it's even more appropriate to push the technology to higher densities if you [want to] have a paradigm breakthrough than if you are just doing the same old thing. When you're not near the limits, people won't listen to you.

EE Times: When you first proposed analog VLSI, you proposed using the material properties of the silicon itself, rather than trying to force the silicon to do what you had already decided to do ahead of time.

Mead: Yes; when the physics gives you an important computational property, then why should you be making that up with some big, complex thing? The physics gives you that for free. That is the underlying point of all this.

EE Times: That almost sounds like making the paradigm match the properties of the material to be used.

Mead: Exactly, which is what the brain does so nicely: It takes whatever is there and makes the best use of it. And if you think about it, that is what our entrepreneurial system does: It takes whatever you can do and makes [the] best use of it.

I firmly believe that as time goes on, people will be making better and better use of the physics, because it is possible. Somebody will figure it out, and they will be able to outrun the competition. Thus, that [approach] will become the next thing that is obvious to everyone.

EE Times: Is that the key to building the smart systems of the future?

Mead: What the nervous system did was evolve to extract the maximum information from many sources of partial information so that the organism could outperform the other organisms around it. Therefore, the others became food for it-which is a lot like [the way things work in] Silicon Valley.

So it's not that we have to slavishly pretend that we have artificial neurons; that is not the point. The point is that, given the physics of the materials that are available, we can arrange systems in various ways-and the one who arranges it in such a way that we get the most out of it, for the least effort in, achieves a superior solution and begins to take over. That has been the history of electronics from the beginning.

EE Times: You have said in the past that we should not be trying to emulate brain functions with chips until we can emulate the brain's sensors-the eyes, ears and so forth. Is that still your approach?

Mead: Yes, because that way at least you know what representation your inputs are using. When you get deeper into the brain, you have no idea what representation is being used.

For instance, there are neurons that activate whenever you see a face with a certain expression on it. And it doesn't matter whose face it is. We have no idea how it does that computation. We know what the neurons' inputs are and we know a little about the first couple of stages of visual processing, but then it gets lost in a morass of stuff that we haven't decoded yet.

We activate specific cells, but that broadens out into a huge array of very concurrent, very distributed computations [about which] we have no clue. We know so little, compared with what we need to know to actually emulate that brain function, that it makes one feel powerless.

EE Times: Don't biologists understand the circuitry of some of the simpler organisms, such as a sea slug?

Mead: I wouldn't say we "understand" anything about the nervous system on the level that we as engineers are used to understanding. In fact, I will say that we don't understand how a single neuron works at the level at which we understand how a transistor works-not even a single neuron. We don't understand how fast the signals propagate, how they get amplified as they go along, how the automatic gain control is done up in the dendrites. We don't know any of that.

For instance, just two years ago some guys in Germany were doing an experiment by putting the same signal in a dendrite close to the neuron and in other dendrites further away from the neuron. And they discovered a funny thing: They found that by the time the signals that were injected further out got to the neuron, they were bigger than the signals that were injected on dendrites closer to the neuron. Doesn't sound like a passive system, does it?

EE Times: It sounds like transconductance.

Mead: That's right! It sounds like there is some amplification. On the other hand, they weren't full action potentials either; they didn't go all the way to the rails. Have you ever tried to build a distributed amplifier, where you could send a signal down a line and it didn't go all the way to the rails?

EE Times: That's not possible, is it?

Mead: Yes it is, but the only way to do it is some adaptive way that senses the signals and makes sure that they don't get too big. So each of a neuron's thousands of dendrites has an exquisite distributed automatic gain control that we have no idea how to emulate with electronics. I kept poking people trying to get them to study this, but they said "Oh, that's too hard," and it is! You're trying to study a submicron phenomenon and you have to get these big probes on it, and the probes load down the dendrite and screw up your results. Believe me, that is a hard way to make a living. These guys are really good, with exquisite technique, but it's just a very, very hard problem.

However, the fact that it's hard to figure out exactly how biology does it doesn't mean that it's hard to figure out how it must be doing it. That isn't nearly as hard.

EE Times: So maybe we don't have to implement biological functions in exactly the same way that biology does?

Mead: We can't do it exactly the same way. So why are we screwing around trying to do it the same way, when we know that in principle it must be doing this kind of thing? Let's get on with doing this kind of thing.

EE Times: That makes it sound like density won't matter in the future.

Mead: What it means is that as densities grow, the issues become more compelling. The problem with worrying about these ideas too soon is that people are [still] able to make the old paradigm work well enough. You need to blow the whistle just before the existing paradigm runs out of gas.

A lot of smart people have spent their lives working on neural paradigms, and they have gotten a lot of insights that will all help with the next paradigm. But it's just really hard to come up with a new idea. I personally have never gotten a new idea; I've always adapted ideas that other people got before, and mostly that's what all innovators do.

EE Times: Still, you seem to see biology as the inspiration for the next new paradigm.

Mead: It's one inspiration. You get inspiration from wherever you can, because there is so little of it.

The nervous system handles [computation] very nicely, though. There is this great cascade of continuous information coming down the dendrites to the neuron, which results in a discrete action potential, which is then distributed out to the axon. Then the whole process starts over again. So we have this cascade of alternative continuous and discrete representations.

We have never even tried to build an electronic system that works like that. It's pretty remarkable that [the nervous system] is an effective organization. We shouldn't slavishly copy that architecture, but the notion that it goes through these alternating representations, plus the fact that there are as many signals feeding back as there are feeding forward, is a pretty remarkable system architecture that we don't have anything like today.

I believe that it is possible to build silicon systems that use a representation like that. We know how to make pulses and send them around, and we know how to measure time delays. It's an extremely powerful representation, and it is my belief that it is a much more powerful representation than anything we are using today.

What this all means is that there is a lot of thinking to be done that we are just not doing today. And we need to be doing it. We need a more powerful representation of information.

We are in a period when whistle-blowers are talking about "atomic walls," but that would only be true if nobody was going to think anymore. And people are going to think. Some of those new ideas will be for wild new technologies that may or may not work out at all. And some of those will be very insightful innovations [with respect to] the nature of silicon itself.

The guys working in silicon today are very smart. The technology is light years ahead of what we had just 10 years ago. People are innovating like crazy today, and if they run into things that look like walls, they are going find ways around them.

EE Times: What about Moore's Law?

Mead: Moore's Law is not a law of physics; it is a statement of human belief. If people believe that it is possible to do something better each year, then they will put the time and energy into it.

Moore's Law is a statement of human nature: When people see hope, they will cause it to come to pass. And today, we have more to hope for than ever before.

Return to: IC @ 40