Experts at the table, part 3: Where are the biggest hurdles, and why the definition of Moore’s Law needs to be expanded well beyond the number of transistors on a piece of silicon.
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. To read part two, click here.
SE: What becomes the most important metric for your customers, particularly as we migrate into new areas such as the Internet of Things?
Kengeri: If you look at whatever is visible in all the analyst reports, IoT is there, but it’s in very fragmented markets. If you look at the inflection point, it goes back to Moore’s Law and cost. It has to be at the right price point. How can you monetize it?
Gianfagna: It’s not a process node that’s important. It’s cost of design, reliability of hitting the schedule, and power. Those three come up in almost every customer engagement. Performance is not that important. The node is interesting but not that relevant. The key is can you get it done reliably, at the power you need and on time? Time is crunched and power is crunched.
Moore: We see similar kinds of things. Our customers are most concerned with schedules for getting their products out. Increasingly, they will take the next node available or the next process when it’s available. It’s not the process that’s driving this anymore. It’s more about, ‘Here’s my product schedule and what’s the best process available for that time?’ Increasingly, it becomes limited by the non-volatile memory. You have to be able to turn things off and keep them off, and you have to be able to fix mistakes. You’re putting in extra circuitry to put the retention in there because Non-Volatile Memory is not scaling. If you think about Moore’s Law, we’re taking the digital part out but we’re leaving the memory part behind. This is about bit cells and non-volatile. The gap between non-volatile memory and SRAM is probably going to slow us down more than just the transistor speed.
Cheng: Power, more than anything else, is the number one concern going forward. It’s not cost and it’s probably not speed. A transaction latency with DRAM in a server application is about 200 nanoseconds. If you can reduce that from 200 nanoseconds to 150 nanoseconds, the power consumption for an entire data center would drop 15%. That’s 15% of a gigawatt, which is a lot of power. The reason that DRAM takes so long to respond is there are a lot of speculative fetches and executions and threads just because there’s nothing to do. They hope they can cancel the transactions when the speculation about what needs to be done turns out to be wrong. If you look at non-volatile memory, the way IoT is going to get deployed is not through battery-powered sensors that get changed every two years because it’s too expensive. Almost all the sensors used in infrastructure have to be able to harvest energy. Once you deploy it, it will stay there nearly forever. Non-volatile memory is a problem because it takes a lot of power. Unless RF energy, solar energy and thermal energy provides enough power for the Cortex M0 plus non-volatile memory, that model doesn’t work. The question is, ‘Can the turn-on voltage for non-volatile memory be lower than Vt?’ That’s the big research. And the last problem isn’t area or size of the transistor. The amount of computing probably can be done in parallelism, so it’s not speed. But it is power. The chip has to harvest energy. Whether it’s data center or IoT or something as futuristic replacing an Alzheimer person’s brain with a chip, power is the number one metric.
Lanza: The minute that the driving force in the industry is health and medical, everything will change.
SE: What you’re talking about here is mixed signal, though, and the analog side has not scaled well.
Rabaey: It really depends on what information you want. I look a lot at the biological world and the brain and how we process information. It’s kind of mixed signal, but it comes down to putting computation at the right place. If you look at the eyes, and the retina itself, it does feature extraction. It doesn’t have to be very good, but you have massive numbers of sensors. You send it to your nerves and brain and expand it.
Lanza: Yes. The analog portion is to get the signal. After that, it can all be digital. Fundamentally, we’re going to have many more phenomena that are not going to have an electronic personality that need to be connected.
SE: Are we talking stacked die?
Lanza: Stacked die will be there. There are many dimensions that will be added to more things.
Kengeri: Sensors will use older technology. You need the digital piece that may require more sophisticated technology. If you want to mix and match all of those, for system-level scaling and 3D, you will need heterogeneous integration. That’s going to become much more important.
Lanza: If you look at all these signals coming in with (IBM’s) Watson times 1 million, you’ll be able to understand everything that was ever said about a disease since it was first described in ancient Greece. It will be analyzed, seen, discussed. It will all be there. All you are doing is bringing new signals that were not there before, and then you have a whole new way of interpreting it.
Rabaey: The question is where you’re going to do that interpretation. So you have every piece of knowledge in the world and you carry all of them with you. You live in a dynamic world, and you want to have as much as possible with you and around you.
Lanza: You’ve defined the infrastructure of a country. This is just the beginning.
SE: So you’re questioning the basic underpinnings of why we’re following Moore’s Law? That would make Moore’s Law a facilitator rather than a goal.
Rabaey: That’s right, because it was a given that we were going to do processors and they were going to have more mass and lower cost. That mindset is still there.
Moore: The promise of Moore’s Law was that you were going to have enough capacity to do the harder tasks that you couldn’t do the last generation. We’ve gotten to the point where we no longer have that promise of capacity—just the raw transistors to do that—so we’re going to have to have a different paradigm. But we still need to grow the throughput at twice the rate to solve bigger problems.
Gianfagna: That’s a more complex vision of Moore’s Law.
Kengeri: It’s helping human civilization progress. It’s not as if anyone is going to slow down.
Lanza: It’s not about how many cycles you have. I always talk about the Internet as a new country, but it’s an international country. It’s everywhere. Someone in Tanzania will be as empowered as someone in Silicon Valley.
SE: Do we have to change out the infrastructure and the tools and the IP?
Lanza: I don’t think there’s a need to change out the infrastructure. But for security reasons, it will be in several different countries. We will need to include changes of what will be needed in the infrastructure. The U.S. will need to have a U.S.-based infrastructure. China will need to have its own.
Moore: Even on a smaller scale, if we’re going to beat the schedule crunch, we need more system design and faster ways to get there. We also need ways to partition the design. If we’re going to go with stacked wafers, especially if you’re going to use wafer-to-wafer bonding, you have to have the partitioning done down to the I/O.
Lanza: We need to see the bigger picture. We are still denying that IP is part of EDA. We need to become part of EDA so we can get the design out on time and into the right areas. We need IP for this. EDA needs to change its name.
Kengeri: Going back to the question, the slowing down of Moore’s Law is putting a lot of pressure on everyone. From a technology point of view, you want to have IP and EDA working together, and in a design that becomes even more important. You need to make the right tradeoffs. So SRAM uses very regular structures. The patterning is all very uniform. You think of logic as random, but there are some very regular structures. You can optimize the heck out of that and get another 5% to 10%. If you can get 5% to 10% area per node, that’s great. It’s changing infrastructure in a different way.
Lanza: Creating limitations in design freedom increases the density.
Gianfagna: Today, how many companies drive the demand and consumption for semiconductors—10 or 12? That’s wrong. Think of the giant companies, take them out of the picture, and see what the world looks like. It’s not pretty. That’s a concentration that feels like a big bang to me. Sooner or later it’s going to explode and something new is going to happen, and that something new is ubiquitous applications that take us in new directions we haven’t seen yet. They have massive parallelism and massive integration and massive network access. That will change the delivery technology and the ecosystem. The applications will change the ecosystem, not the other way around. Delivered throughput for less money is fundamental to this, because people aren’t going to pay more money. They’re going to pay the same or less. But what Moore’s Law means is going to change.
Moore: Moore’s Law needs to continue at an electronic systems level, not a transistor level.
Lanza: We need to be applying the intellectual energy we have to make the design of systems optimal. We want to make sure we have the ability to make those new systems take off exponentially as quickly as possible. That’s what’s going to change.
Cheng: There’s a lot to be done yet in Moore’s Law. There’s still a lot of innovation to be had. Power is the number one issue. Everyone says a lot of eloquent things about how to move Moore’s Law forward, but there are still a lot of gaps. There is a lot of hope for the next two innovations, and after that, who knows?
Kengeri: It’s definitely slowing down, but there are innovations in the pipeline that hopefully will continue Moore’s Law. System-level scaling is going to be really important.