Moores Law extended: The return of cleverness
07/01/1997
Moore`s Law extended: The return of cleverness
An exclusive interview with Gordon Moore
Ed: How long will the industry continue along the curve of Moore`s Law? When will the slope change again?
Dr. Moore: The slope has changed a bit already. The original version of Moore`s Law was doubling the number of transistors on a chip every year, and then I said it`d be slowing down and doubling every two years, and we`re pretty near that now. For a while we were doubling every 18 months. It didn`t roll off quite as fast as I thought it was going to. If we can stay on the SIA Roadmap, we can essentially stay on the curve. It really becomes a question of putting the track ahead of the train to stay on plan, which is kind of one generation every three years or cutting the minimum dimension down to 0.7 of the previous generation.
Quarter micron will be here before very long, followed by 0.18. All this looks reasonably do-able as given in the Roadmap without any real changes. It would be nice to have a shorter wavelength lithography system, but we can squeeze 0.18 out of the machines we now have if required.
Below that, we really depend on having a shorter wavelength. For example, 0.13 micron in my view is hard to do with
193 nm, but the guys that have to really make it work are not as concerned as I am. By using all the tricks, phase-shift masks or whatever, they think they can get it to work with a high numerical aperture system. So if we grant them that, the real problem in my view comes in the next step - to go down to 0.10 micron. There you need a shorter wavelength; and lenses get opaque below 193 nm. There just are no optical materials left. There`s hardly anything left at 193. So we finally get to the point when we have to do something other than optical lithography, in the classical sense. And to me that is one of the most significant barriers we have to get over to stay on this curve. Of course there`s been a lot of work going on in this area. A lot of people have been working on x-rays....
Ed: EUV, too.
Dr. Moore: I usually describe extreme ultraviolet (EUV) as an "equally challenging approach." Regarding x-ray lithography, a lot of money`s been spent on it. It was initially going to be the technology for half micron, but it`s gotten pushed further and further away. The problem it has - the need to make masks that are thick enough to absorb the x-rays - gets increasingly difficult as you draw narrower and narrower lines. So to me, conventional x-ray doesn`t seem to be advancing as fast as the requirements for it are advancing.
Electron beam can do the job, but it has been horrendously slow and as we go smaller, the amount of information you have to put in a given area goes up as the square of the minimum dimension. So it just puts an increasing problem on the throughput. The Japanese are doing some work there to use other than a point beam, shaped somehow, to speed up the writing. EUV, to me, is the most attractive possibility, probably because we know the least about it!
Ed: We haven`t identified all the potential problems?
Dr. Moore: Yes, well that`s it. At one time x-rays seemed attractive, too, but you end up needing a synchrotron and then the mask problem reared its ugly head. EUV requires that you make better optical surfaces than have ever been made previously. The light source is not really a proven thing yet. You can get light in that wavelength range, but can you do it reproducibly and reliably over a long period of time? If you take a generation as every three years, 0.18, 0.13, we`ve got nine years to put it into production, but that means in half that time we`ll need a reasonably acceptable system. So to me, lithography is the thing that we really have to be concerned with.
Ed: There`re a number of initiatives going on right now to try to address the mask problem for x-ray. As optical is getting pushed further and further using optical proximity correction and phase shifting, just as you mentioned, it gets progressively more difficult and progressively more expensive. We`ve seen studies that claim that for some critical layers at the 0.25-micron generation, x-ray may be more cost-effective in high volumes than pushing DUV.
Dr. Moore: I have trouble with it at the 0.25 level. At 0.13, I don`t know, that`s iffy. I guess you`re getting near the end of the optical stuff, and your costs are clearly going to go up there. If we have to use 193 nm at 0.13, that`s going to be pretty difficult and the mask costs will go up. We need a solution beyond 0.13, anyhow. Intel is supporting EUV very strongly. We think that EUV has the best chance of being the kind of system that we`ll want to use.
Ed: If we are starting to approach the practical limits of 2-D resolution, does it imply a move to 3-D transistor architectures, using some sort of MBE or other controllable layering process to do junction formation in a vertical direction, or other fundamentally different transistor architectures?
Dr. Moore: It`s not clear to me that you buy yourself a lot there, though in some specific cases you can. In some respects it`s done already, for example, the active loads in SRAM cells use thin-film transistors. I think there are simple applications like that, but I have a tough time envisioning how we get to multilayers of high-quality transistors.
Ed: You mentioned that interconnects are already a problem for greater device density. One burning issue for the end of the 1990s is what will be used both for metal and for dielectric in another one or two generations. Mark Bohr, in Intel`s Portland Technology Group, did some very nice modeling of RC delay improvements in going from aluminum to copper.
Dr. Moore: Copper is an intractable material. The reason we don`t use copper is not because we haven`t tried over the years. Getting bulk conductivity in copper in the kind of thin-film structures we`re talking about is difficult, and it`s awfully easy to make the copper resistivity higher than aluminum.
 |
Figure 1. a) Approximate component count for complex integrated circuits vs. year of introduction; b) decomposition of the complexity curve into various components; and c) projection of the complexity curve reflecting the limit on increased density through invention.
Ed: It looks like people are already fairly convinced that they`re going to try to use copper anyway. But again it does seem to speak to reaching some fundamental limits. What`s beyond copper other than optical interconnects or superconductors?
Dr. Moore: Nothing obvious. When high-temperature superconductors first came along, we were all excited about the possibility. The more we learned about them, the less excited we got. They don`t offer anything like a room temperature superconductor now. People who have looked at small devices suggest that at around 0.05-micron minimum dimensions you get into some fundamental device-physics problems where transistors don`t behave in a way that you can make ordinary circuits with them anymore. I guess if you take the rate of progress that`s projected, we get to the 0.05 micron in something like 2017, I think one of our guys said. So that`s the end of Moore`s Law! Somehow I don`t think we`re going to go "Klunk!" into it; we`ll probably have a somewhat smoother landing than that.
Ed: As you originally phrased the law, you were merely identifying that the number of devices/chip was increasing. Now of course that includes chips getting larger and device structures getting smaller. Also, structures have shrunk considerably and chips haven`t become orders of magnitude larger.
Dr. Moore: They`ve grown a lot though. I`ve looked at die size changes from the original transistor. Dice shrank down a bit at first. But from the minimum die size up to something like a Pentium Pro is about a 2500 factor. So there`s a lot more area we`re dealing with now.
Ed: One aspect of the increasing density has been increasing speed and thus performance. But if you merely put more structures on larger chips that run at the same speed, it`s a minimal advantage.
Dr. Moore: I don`t think so. What`s driven the industry is lower cost. The cost of electronics has gone down over a million-fold in this time period, probably ten million-fold, actually. While these other things are important, to me the cost is what has made the technology pervasive. I`ll hate to get to the point where we don`t get the speed improvements also, but that may happen for such things as power limitations going forward. If you keep scaling, power becomes an increasingly important problem. We`ve been able to fight it so far by lowering voltages. At somewhere below a volt, I don`t know where, that approach runs out of gas.
Ed: Let me step back a bit. You mentioned that what has primarily driven the industry is cost/function. Of course, speed increases and size decreases are all wonderful benefits. But if we may be reaching some fundamental material limits in lithography and interconnects, does that imply that the industry might continue on a Moore`s Law curve with a much shallower slope?
Dr. Moore: We`ll still continue to evolve, yes. The update I did on so-called Moore`s Law in 1975 at the IEDM meeting (Fig. 1a) resolved where the improvements had come from previously (Fig. 1b). As I recall, the one factor was obviously higher density, and one was making bigger die. The third one was we were learning to pack things more efficiently, what I called "cleverness." We went to isolationless structures, for example.
At that point, in 1975, we`d come to where the most complex device was a CCD memory that had active areas packed close together. So one of my principal points in the changing of the slope was there`s no more room for cleverness. We`d removed all the dead space. So I predicted the curve was going to change (Fig. 1c).
I think we`re looking at the same kind of thing again. If we say we can`t improve the density anymore because we run up against all these limitations, then we lose that factor and we`re left with increasing the die size. So the slope changes; it probably makes transistors on a chip double every four or five years. By that time we`ll be at the point where you can put a billion transistors on a logic circuit (Fig. 2). So you have the flexibility of what do you do with a billion transistors worth of circuitry. The slope will have changed again on the rate we should be progressing, but the industry is certainly not going to stop.
Ed: In this eventual shift, would improvements be driven more by cleverness?
Dr. Moore: Yes, the return of cleverness.
Ed: If this is the case, if the greatest strides might be made by a return to cleverness instead of the more brute-force method of just shrinking everything, might companies begin to differentiate themselves through extreme cleverness?
Dr. Moore: I think it`s a realistic way for a fair portion of the industry to go. It could change the economics considerably.
Ed: Over the last 10-15 years there has been a movement for semiconductor equipment companies to share in the burden of process development. Twenty years ago, Intel may have just bought a nice piece of equipment from somewhere and developed the process entirely on its own. Today, the equipment companies are selling complete turnkey processes that you can buy and then customize a little bit for your particular process. Some people have suggested that equipment companies should eventually take sole responsibility for process.
 |
Figure 2. Moore`s Law projected out 15 years predicts a) microprocessors with 1 billion transistors will b) operate at 10 GHz.
Dr. Moore: I think that model is wrought with real problems. The integration of a 200-major-step process is also an important part of the whole thing. Somebody has to do that and that`s something the equipment companies are not currently really able to do.
Ed: For Intel`s success today, what would you say was the relative importance of design vs. process?
Dr. Moore: It depends a little on product area. In the old memory days, the process end was probably more important than design. With current microprocessors, design is certainly critical, but product cycles are a lot shorter and the product`s success is very dependent upon process capabilities. So we consider the two of them to be equally important. In fact, they interact a lot.
Ed: How do the foundry business and the fabless companies fit in?
Dr. Moore: We can be more aggressive in designing our products because we know exactly what technology is being developed to run them. This is significantly different from doing foundry designs where you`ve got to design for a least-common-denominator process technology. You can`t push the designs nearly as hard if you`re going to run on a foundry`s processes. Similarly, the foundries can`t tune their technologies as much because they don`t know exactly what they`re going to have to build. So, for many of the products that we make, having design and process together is really necessary.
However, the fabless semiconductor companies have been reasonably successful, a lot more successful than I thought they`d be in the beginning. Similarly, I thought the foundry business was the worst investment imaginable, and some foundry companies have done pretty well. TSMC has made a very nice, profitable business out of it. We use them a lot!
You can split the industry in a variety of ways. The big systems companies, at one time, all thought they had to have semiconductor capability in house. They subsequently found out how expensive it is to keep it at the state-of-the-art level. Most companies that got in later decided that they could avoid it and buy semiconductor products from companies like us. I think the various companies will find different places where they will want to split between internal and external. It depends a lot on the kind of products they`re making - what they consider are their advantages.
Ed: Is there any correlation between this and the return of cleverness? If you have greater cleverness in design, does that necessarily imply a need for a more closely synched and custom-tuned process flow?
Dr. Moore: It`s more of a question of how hard you`re trying to squeeze performance, and how close you`re working to the limits of what the technology can do. The closer you are to the limits, the closer you have to be integrated. If you can back off a generation in complexity or performance, then that integration doesn`t need to be nearly as close. We`re in a lot of product areas where the most advanced technology is not necessarily an advantage. The performance isn`t needed in the application, so you might as well run the old processes and the depreciated equipment.