Experts at the table, part 2: The original intent of Moore’s Law; new variables for the equation, including cost per function and cost per user experience; why some costs won’t go down anymore; vertical versus horizontal design.
Semiconductor Engineering sat down to discuss the future of Moore’s Law with Jan Rabaey, Donald O. Pederson distinguished professor at University of California at Berkeley; Lucio Lanza, managing director of Lanza techVentures; Subramani Kengeri, vice president of advanced technology architecture at GlobalFoundries; Charlie Cheng, CEO of Kilopass Technology; Mike Gianfagna, vice president of marketing at eSilicon; and Ron Moore, vice president of marketing for the physical IP division at ARM. What follows are excerpts of that conversation. To read part one, click here.
SE: What’s changed with Moore’s Law?
Lanza: The title of Moore’s article was “Cramming more components onto integrated circuits.” Gordon Moore told me the article he published was a marketing article because the transistors and chips did not have the same level of reliability. We wanted people to understand what the future would be. But one thing you need to be careful about is the key variable in guiding the evolution of Moore’s Law has changed. At the time we discussed what the key variable should be, and we decided it should be the frequency or speed. That was literally out of a meeting of four people. At a certain point we need to understand what is important, whether it’s power, performance, reliability or security—it’s going to be different than in the past. One thing that is changing significantly is that the competition in the semiconductor industry has moved from companies to countries, if you look at the cost of the fabs. On top of that, you will have an incredible number of options. Size is not necessarily the key variable. We have to think about the systems we are building, not the chips we are building.
Rabaey: It’s a lot harder when you say, ‘What is this going to be used for?’ You have to engage to do that, and the semiconductor community has not been very good at that. Information technology is changing every single profession.
Lanza: This is not about design automation anymore. It’s about something else.
SE: There are companies such as Apple and Google taking things in-house and focusing less on the speed of the processor than what the system can do. How does that change things?
Moore: They’re doing that at the systems level, too. Japanese companies are organizing around the infotainment of the car. They’re getting all the specialties there. Other companies are organizing around medical systems. The idea is that there will be some high-dollar things that do provide twice the system at half the cost. But there will be some things we’re used to seeing cheaper every generation, like our cell phones, and we may just have to pay the bill for that. Luxury items may not be cheaper, but there will be innovation on big-ticket items that are big volume.
Kengeri: This may all be looked at from a system-level scaling point of view. The new metrics may be cost per function or per user experience. I’m not sure how to measure that, but the focus will be on system-level scaling. It may involve interposers or 3D packaging. At that level, you have some leverage to continue Moore’s Law for several more generations.
Gianfagna: It takes a different form. You can chase feature sizes and switching speed all you want, but it’s the delivered throughput for the application. Moore’s Law takes on different meanings, but improving the performance per unit cost is relevant forever.
Moore: Yes, that’s correct. It’s not about manufacturing anymore. It is about the entire ecosystem.
Gianfagna: The concept of higher complexity for a given cost continues.
Moore: And we left it in the manufacturing so long largely because we’ve been waiting for EUV.
Kengeri: It’s an entire ecosystem that’s involved here.
Gianfagna: Yes, it’s a more complex delivery today with materials and manufacturing equipment, and it gets even more complex in the future.
Rabaey: We haven’t figured it out yet, though. Why do all these companies go for vertical solutions? We have failed utterly at horizontal system design. It’s still an art. If you want to make that happen, you have to deal with every part of it. That’s what all the big companies do. They have the whole chain. They have some power savings room at the bottom and they take advantage of it. But the system-level design thing has been a failure.
Cheng: Our customers are all talking about the same thing—better delivery, more parallelism. But it’s not saying semiconductors will not deliver more value and the value has to come from verticalization and different software. Those are certainly important, and it’s not as if we can’t get more value from semiconductors because we can’t do EUV. Customers are coming back to us and asking where the next innovation is going to come from. The problem is that devices fundamentally have not changed, so we are pretty focused on changing the devices. If something doesn’t change for 30 years, you have to believe there is something better out there. The first step is non-CMOS. People haven’t looked at it seriously because the risk is too high. If you’re building a $5 billion fab you don’t really care if CMOS isn’t that good. Your big concern is that you don’t want to take a chance.
Rabaey: I agree partially. Everyone has been looking at devices, but they’re all looking for the same thing, which is the next ideal switch. They’re so stuck in the mindset of digital logic for so many years that they can’t see beyond it.
Cheng: I disagree. I went to one of the largest semiconductor manufacturers and we proposed a new device. There were 30 engineers there, and not one of them has studied this stuff in school. The reaction is always the same—a young engineer has an idea for a new device. They all say, ‘What are you crazy? Their competitor has 500 engineers working on something else so that must be right.’ But only the fab manager meets with the chairman every day, and that’s because he has a $500 million P.O. that has to be signed. The design teams have no influence at all.
Kengeri: We spend more than $1 billion on research every year, to go through all the exploratory phases and to experiment on real silicon. Researchers across the semiconductor industry are always taking risks because innovation pushes the envelope. We do pathfinding, we shortlist critical elements, and then we come back and revisit from a SoC point of view. At the same time, we are focused on business and can’t take our eyes off of that, but we do explore and keep our options open. That’s why we have a number of companies in our R&D ecosystem sharing the cost and we will continue with that model. At this point, we are exploring the next two nodes. The real issue is not just the devices, though. If you take the conventional switches, they work, but the real issue is the back end RC. There has not been enough innovation on the back end of line. You may have the best switch, but without the right interconnect, SoCs cannot extract full value of the device, and that’s really a big challenge. But going back to an earlier point, I don’t agree that verticals are necessarily the best way to solve this. If you look at ARM, which is enabling a big portion of the industry, the whole system—the software to manufacturing—all of that has been working really well, in an open ecosystem. If you want to change the power management scheme at the system level, the entire ecosystem innovates and supports the end goal.
Lanza: Today we have a system that is not working well. It’s a system of virtual IDMs. But that’s an issue with money. It’s not an issue about how you design. We need to change the dimension completely. If we keep thinking the future is going to be exciting because of how many transistors you can have or how many chips you can make, we are in trouble. The smartphone came out and changed things completely. But this market is now slowing down. The biggest change here is that people are more mobile and you can do things any time of day or night. But the big change that is coming is the Internet of Things. That will change who is providing the direction of the evolution of the future. The IoT is a very simple thing. It’s a new country of things, and it’s going to create certain demands and supplies, particularly in health care and medical.
Rabaey: There’s no question about it. The correct mindset is where is this leading? And you can see this evolution happening from set-top boxes and mobile phones, wearable devices, immersible devices. It’s very foundational. We’re going to have a whole world of information—the cyberworld—living by itself. And on the other side you have this physical world where you can build stuff, as well. What we’re doing now is building a high-bandwidth connection between the physical world and cyberspace. It’s going to change whatever task we do. The right question is, ‘What do we need to make it happen?’ What do you need to build to make it happen? If you look at how robots are operating these days, they’re all based on machine learning and stochastic computing. What building blocks do you need? You have to start from the application and work backward.