The Impact of Integrated Circuits, lecture by Robert Noyce

Thanks very much, Gwen. First of all, I hope there's some money left over for dinner and for the museum. I trust there will be. It's interesting to me to see what an enormous amount of publicity the microchip has gotten recently, including that article by Tom Wolfe. Incidentally, on that article, I do want to point out that Tom Wolfe is first of all a novelist and only secondarily an historian. But everything that appears in print, of course, is true, as you've all learned, particularly if you ever read something that you know something about. Or perhaps after the last issue of the Proceedings of the IEEE, which was on personal computers, the editors, Amar and Gupta and Tong, were comparing the progress in the computer and the aircraft industries. They pointed out that if in the last 20 years the aircraft industry had made the same progress that the computer industry had, a 767 would cost $500 and go around the world in 20 minutes on 5 gallons of gas. In 1959, the Cadillac would cost you a quarter. You wouldn't bother to pay a parking fee because you just put it in your pocket after you got where you wanted to go. Well, it has been a lot of progress. Certainly the integrated circuit and the computer is now affecting every part of our activity. One of the things that I like to think of is that we as a society may be progressing in the same way that medieval education progressed. You may have some very interesting things that were used in there: arithmetic and geometry obviously, but also finally astronomy, which is a good way to see order in the universe, and then music. Maybe there's a great deal of correlation between what is happening in the computer field and what that advanced education was in medieval times. Since this is a museum, I think it's maybe fair of me to be somewhat retrospective. As I was driving in tonight, I was thinking that the integrated circuit is 25 years old, so maybe we are indeed reaching middle age. It's interesting that only recently has it gotten a lot of attention. I often thought that when the real fun was going on was back when we were making all of these vast changes. The reason it didn't get any attention though was because it was a tiny little business. This is a slide that I used in 1960, sort of historically, listing what the semiconductor market was at that time: $3580, $140, $210, $360, $550 by 1960. Indeed, nearly half of that is two-terminal devices, about half of it is transistors. Silicon was taking a relatively small share of it; most of it was germanium. What was going on in that decade of the 50s? Those of you who were lucky enough to be here last year heard about from Les Hogan as everyone was trying to figure out new and better ways of making transistors. At one of the solid state conferences, there was a little box, and in this little box is a little tiny cube of germanium and three pieces of wire. If you ever got lost, what you did was to take the little piece of germanium and start making a point contact transistor out of it, whereupon 10 people would lean over your shoulder and say, 'That's not the way to do it.' Then you would turn around and say, 'Well, where am I?' There were lots of things like that. You would take a piece of semiconductor germanium in this case, put indium on top of it, and melt it just enough so that the indium would dissolve some of the germanium and then recrystallize, leaving that germanium doped with the p-type dopant that indium was. You did that on both sides and you make a PNP transistor. Well, while that was going on, there was a great deal of research going on too, trying to understand various properties about germanium. One of the things that happened was when you heated germanium up and then cooled it down, it changed from n to p type, and nobody could understand. At the same time, of course, we'd been making these transistors with the n-type germanium with the indium in contact with it. It was a couple of years before people realized that the indium acted as a getter to pick up all of the impurities that were left on the surface of the germanium, mostly copper, so that it didn't convert the germanium. Anyway, all different kinds of ways of making transistors: alloy transistors, grown junction transistors. Can you imagine making a single bar of semiconductor, pulling it from the melt, doping it one way and then the other way, over-doping so that you have a layer slowly grown as thin as possible? Now about the best you could do was a range. It was in the 1954-55 era that we started worrying about diffusion as a way of getting impurities into these semiconductors, and that gave us good control of the depth dimensions. Things soak into a solid very very slowly, so that gave us really the control of that one critical dimension. Then the other question was to get control of the other dimensions. I think that some of the first work that was done on that was done at Philco because they had been working on this in the right across the hall from the laboratory that was working on etching shadow mask tubes for color TV at that time. So slowly, photoengraving replaced what had been used as a way of making diffused transistors, namely making a slot in a mask and then having one dopant evaporate through the mask that way to a plate, the other one through a mask that way to a plate, so they were closely spaced but separate. One would be a p-type dopant, one would be an n-type dopant. That was the way. Oh, there were all sorts of clever things that people were doing. Anyway, the photoengraving worked out a lot better. Then the other critical element that was to set the stage for the integrated circuit was the invention of the planar transistor by Jean Hoerni. Now the reason the planar transistor was important was that impurities on the surface of the transistors at the junctions of the transistors could actually have more impurities on the surface than you had in this entire crystal of the semiconductor, and it sort of bloused up the characteristics of the transistor, let me put it that way. So Jean's idea was to leave the silicon dioxide, which was a very good insulator, on top of the transistor when it was being diffused, just leave it there. It was a good insulator and protected the transistor, and indeed it did. There were some other things. The Air Force was working on things like Project Tinkertoy; they were trying to get all kinds of elements into a nice little square package so that you could stack them up and make nice square little modules out of them, so that maybe you could put them in missiles or something like that. There was the concept of molecular engineering, the idea that you could create an electronic function by putting down a molecule of this and a molecule of that, sort of like maybe painting a picture by adding paint to the canvas until it created the impression that you wanted. Those didn't really work very well. There was an awful lot of hand labor in making transistors. Think of it this way: we took a square inch of silicon, we were making transistor chips on those that were about 10 mils on a side, so that square inch would have something of the order of 10,000 transistors. Now on each of those 10,000 transistors, I had to attach a couple of wires as well as soldering it down. Well, that was the day when the assembly plants' employment outnumbered the rest of the company by two to one. Much of that assembly was done by hand. I remember when I was in school, I'd hand in my paper and the professor would come back with these big red markings on it saying 'hard way,' and then he'd jot down a couple of sentences which clearly made it much easier to do by using some other method entirely. I guess that stuck with me. One of the characteristics of an inventor, I think, is that he's lazy, that he doesn't like to do it the hard way. Well, putting those 20,000 wires on meant something else. But the thought of printing a circuit on top of those transistors instead of taking the transistors apart and putting them in the printed circuit basically was the genesis of the idea of the integrated circuit. So you can see the elements coming together: the photoengraving so that you knew exactly where everything was on the wafer, it could be reproduced exactly; the planar transistor which left a perfect insulating layer on top of the silicon so that you could indeed put conductors on top of it which would not short out. We had to get some substitute for taking all of these little chips apart so that they would be electrically isolated. Well, there were three ideas that popped up at that time. One was junction isolation, which was you put a back-biased diode basically between each one of these things, or a couple of them pointed in opposite directions so the current couldn't go that way or that way. Indeed, I put that in the patent application, but it turned out that Kurt Lehovec had thought of that years before when he was at Sprague. Another technique that was thought of by Jay Lathrop at Fairchild at the time was to etch them apart. The patent I think was to use intrinsic isolation, that is to convert all of the silicon in between things to an insulator. I got the patent on that, but it didn't work well anyway because the way to do that is to neutron bombard or gold dope it, and that means the lifetimes are too short and there's a lot of leakage current and that sort of thing. So junction isolation was the one that is being used broadly now. Then there was the question of getting the other circuit elements that typically are in those early DEC modules, like resistors and capacitors and so forth. Well, those were relatively easy; you can make resistors and capacitors in the same silicon. But still, after the original concept, things were very very slow to develop. In a sense, the idea was there at the time, but it was not practical to do anything. It's much like the junction transistor took two years after Shockley described the junction transistor before the first one was built. Part of it was the tyranny of numbers. We were still not 100% yield on transistors at that time. Let's say that it was 60% yield, I'll make it easy, 50% yield. Let's say then we wanted to put 10 transistors together. Well, you got a yield of 1 over 2 to the 10th. Early integrated circuits were still very slow. We were using simple logic forms, saturating logic, and they were indeed slow. As a matter of fact, if we look at that last picture, the one of the first of the DCTL logic chips is that one in the upper right hand corner. From there on, we go to about a two-year spread down to the lower left, to the lower right, and to the middle. As you can see, we are making steady progress. The other thing was that there was a lot of market opposition to using these integrated circuits. 'How can I possibly use your circuit design? My specialty is designing computer systems. How could I possibly use something that somebody else designed instead? I lose all of my proprietary position.' The other piece of that was that it was a time when we did worst-case design, that is, we took every element, tested it to the worst possible environment it could be in, and assured that it would work in that environment, and those were the only units we used. Now in the integrated circuit, the transistor in the middle of an integrated circuit is not going into another electrical environment; it is going to have exactly the same transistor surround it in every case. So we made progress. This is sort of the classic Moore's curve, that every year for at least a while you could get something twice as complex as you could the year before. That extrapolates, of course, to a million elements in 1980. It didn't quite make that unless you're going to allow for the introduction of new things like magnetic bubbles, where indeed you had million-bit bubbles by that time with no trouble. Also, we had to change from bipolar to MOS up there to the newer technologies. The cost of leaves that you had to attach to it, those 20,000 leads per square inch of silicon that we were talking about earlier, that was one cost element. The other cost element was the decreasing yield as we go to higher and higher complexity, where the cost then gets to be proportional to some exponential in the number of elements that is in the chip. The combination of those two costs will have a minimum somewhere, and that's all I'm trying to show on this slide. Incidentally, the assembly and test costs are not proportional to n, they're proportional to 1 over n. That would move down and to the right, and indeed that is why we've moved up that complexity curve. It was done through lots of different ways. One is we used larger and larger wafers. Believe it or not, back there in 1963 we were using 5/8 inch wafers, and in 1965 that was a 1 and 1/2 inch wafer. By 1970 we got up to about a 2-inch wafer. I should have had dimes on those others for reference. The other thing that we were doing was to increase the die size. The straight dog work cleaning up the dust in the production machinery. The other thing that you're very familiar with is the drive to make things smaller. It has two effects: one is to get more on a given density, a given area of silicon, but it has another salutary effect too, and that is roughly that the speed of these circuits is proportional, or the delay let me say is proportional to their dimensions. Speed is inversely proportional, and as we are getting things smaller, we are now talking about in production two-micron circuits, talking about one micron and 7/10 of a micron circuit. So we're pretty well following this curve still, which was drawn in 1978 as I recall. But those are small dimensions compared to some others that you might like to compare it with. If we look at this one, there's an index mark right here which is 10 microns. That much is 10 microns on the screen. So what is that? That's a neuron, that's what making this speech synthesizer work. Two things that came in during this period that were important: one was MOS versus the bipolar, the second was epitaxy. Prior to the use of epitaxy, we had only been able to make the surface more impure than the underlying material. Now it was possible to have impure underlying material, high conductivity underlying material, and put pure semiconductor on top of it. So it was another great bag of tricks. The first set of integrated circuits were really pretty easy; they were straight Boolean functions. The designers had a habit of having lots of leads come out of a circuit, and that was something that the semiconductor people just didn't like at all. There were arguments ad nauseam as to what was the optimum ratio of leads to gates. The one I always like to use was that it was simple at both ends: a diode had two wires to it, and so did your TV set or your computer. But the designers weren't buying that. The other problem that we ran into was that as we made things, we were talking about making this great huge thing, let's say the whole computer, and only make one of them because as you know, von Neumann suggested that four computers would saturate the world's need. We were talking about making things in the millions, and the computer companies were thinking of making computers in tens of thousands per year. It just didn't seem to be any reason why we would want to make 10,000 or something. Finally, it would be just the cost of design; there wouldn't be any manufacturing left, and we were still thinking of ourselves as manufacturers at that time. There was one element: we decided really that memory was the thing to concentrate on. This is a picture of a large core, unfortunately it's 15 mil core, on top of the first of the static semiconductor memories, the 1101. I think it's obvious just looking at that that what's under the core should be cheaper to make than the core. It wasn't, of course, but it should have been. And finally, it became so. It was very rapid cost reduction in core. This is a slide I made up in 1970, I believe, no 71 I guess, back when Intel was just a little company, so you do your own slides. By the time we got something a little bit more complex, they became cheaper, and new ideas were coming in. The way you made them cheaper was to make them more complex, same function. There were 16 kilobits of RAM over three generations, each about a couple of years apart. This was a slide prepared in 1974 projecting at the time when semiconductor penetration was about 30% of the memory market, projecting that it would go to 80% in 1977, and indeed it did, and now it's about 100%. In the early days, there then the next one that came along was motivated by the fact that having gotten the cost of memory down and really dependent on semiconductors, it looked as if it was then time to go back and do something more about the logic. This is the story, if you will, of the microcomputer, microprocessor. I don't know quite which is which in common parlance now; they've been used interchangeably too long. The idea was to take a system and integrate that into just a few chips. The problem with that was that if you looked at the total variety of various printed circuit boards that were being made at the time, there were thousands of them, and the cost of reducing one of those things to a circuit design was in the tens or even hundreds of thousands of dollars. As a matter of fact, it is still almost true that the design of a circuit really depends still on the number of transistors in it and is almost linear in it. Now you can say that's pretty terrible performance considering the fact that you have computers to help you, but we needed then to find a way to take advantage of this new technique. We needed to find a way of finding large applications for a relatively few different circuits, and indeed that is what the microprocessor allowed us to do. I think it was indeed a new era of integrated electronics. Early on, they were used for calculators and simple controllers. But I recall going around talking to people who wanted me to do their custom design, and at the time I said, 'Well, could you use these?' And slowly the idea took on. Now it was not building computers at that time, but rather it was a simple controller that could be programmed to give you a given logic function when you switch the thumbwheel to the 767, you're listening to channel 10 instead of channel 12, you know, little control thing. Indeed, that was one of the early applications, not on the 767 but on the 747. Then of course this spread to the world, and we are now indeed moving on to the era of getting literally hundreds of thousands of transistors in those processors in that little chip of silicon. I might point out that the 6600, which was a very large scale machine that Seymour Cray designed, we'll see pieces of it in the museum finally I understand, had only half a million transistors in it, and that was a supercomputer. That will be available within one chip within a couple of years. So we're following behind at 15 to 18 years from the supercomputer to the chip. Some trends: we'll move to CMOS instead of NMOS just simply because it's easier to design with low power dissipation on that piece of silicon; it gets too hot otherwise. It's easier to what you do with NMOS design now is to see how fast a particular node has to be and then tailor it to be that fast and use as little energy as possible. Well, with CMOS you don't even bother, just make them all alike. Bipolar will probably no longer have an advantage over CMOS as the dimensions are controlled. It used to be that the diffusion dimension, the depth dimension, was much more accurately controlled than the XY dimensions. Where is the limit? It's quite a ways away. I think the fundamental limit that we're facing, and there are ways around this one too, is that if we're going to make computer-like equipment, we'd like to be able to tell the difference between a zero and a one, and to do that we have to have signal levels that are large compared to thermal noise. At room temperature, we're at about 100 or 200 times the level of noise. The power required for operating a microcircuit goes down with its dimensions. So if we can go down say another factor of 10 in the linear dimensions, that means we can go up a factor of 100 that we can pack in the same area. And indeed, if we continue the trend that we've had in the past, we can go up easily another factor of 10 in the area that we use. That means that instead of the 100,000 to million transistors that we put on a chip these days, within, oh, say a decade, it would be possible to put between 100 million and a billion transistors on that same chip. I think of the early prognosticators who said that four computers would saturate the world's need for it. When we say we can't think of anything to do with a billion transistors which you can reproduce at will, I think that the sky is still the limit, and that it really depends on what we can imagine to use this marvelous new tool of electronics for. Thanks very much.