Moore’s Law at 60: how it’s still changing the world
53:57 Share

Welcome to today's episode of Lexicon. I'm Christopher McFadden, contributing writer for Interesting Engineering. In this episode, we dive into the future of computing with Sanjay Natarajan, SVP and GM of Intel Foundry Technology Research. From keeping Moore's Law alive to breakthroughs in advanced packaging AI materials, discover how Intel is shaping the next era of innovation in semiconductors and global technology leadership.

But before getting into today's episode, make sure to check out our latest merch at Interesting Engineering Shop. Engineer your style with our t-shirts, mugs, and discover new products. Now let's continue with today's episode. Sanjay, thanks for joining us. How are you today? I'm very good. Thanks for having me. Our pleasure, of course. For our audience's benefit, can you just tell us a little bit about yourself, please? Sure. My name is Sanjay Natarajan. I'm a Senior Vice President at Intel Corporation. I'm

I am the general manager of Intel's technology research group, which is doing kind of all of the advanced research in support of forwarding Moore's law, as well as some important adjacent technologies like quantum computing, gallium nitride, a few other things that are going to be critical for the future of the semiconductor industry.

Fantastic. So on the subject of Moore's Law, it has been the cornerstone of technological advancement for I think nearly 60 years now. What do you see as the most profound ways it has shaped the modern world and how do you think it remains relevant today? Yeah, let me give you a few examples of how I think it's shaped the modern world and why I think this is still a relevant topic today. If you don't mind, I'm going to, before I do that, just kind of catch your audience up on some Moore's Law basics.

that might be helpful. So what we're talking about is came from a paper that Gordon Moore wrote back in 1985, I'm sorry, 1965.

Gordon Moore was one of the co-founders of Intel, and he wrote a really short, important four-page paper back in 1965. And the paper had two graphs in it. I'll describe them both for your audience. The one that I'm going to describe first is the one that everyone talks about when they talk about Moore's Law. The other one that nobody talks about is the actually more important graph in the paper. But it never gets its due. So the one graph he talks about first is

The number of components on a chip, he plotted four data points across a few years, drew a line through it and said, looks like it's doubling every year. And then a few years later, he added a little more data to it. A few years later, a little more data. In the early 70s, he said, it looks like it's doubling every two years. And the number of components on a chip became synonymous with the number of transistors on a chip. And that became known as Moore's Law.

Side note, Gordon Moore never liked the phrase Moore's Law, the term Moore's Law. I had the pleasure, the honor of being able to talk to him personally about this early in my career. And he confirmed that this wasn't his idea. He wished it was never named after him.

I believe named the Carver meat at Caltech and the name duck. And that's really kind of what the genesis of Moore's law was back in the 1960s. The other graph I want to mention is the graph that says these new technologies will get cheaper.

couple of years. And that was really the point. The point was, and that, you know, one of the little secrets Moore's law, the paper came out in an electronics journal, but it was really an economics paper, you know, disguised as an electronics paper. It was really the point of his, of his observation was how the cost of these components would come down over time. And I think that's, that's been a little lost in the transistors per square millimeter view of it, but

but it really is a cost argument. Now, fast forward, we've basically been on that treadmill for on the order of 60 years. And if in fact you do plot the number of transistors on a chip over that 60 years, it's been pretty exactly, you know, two X every two years. So it's a pretty remarkable observation that's held true even up until now.

The aspects of it that he never talked about that we sort of take for granted when we talk about Moore's law is that these transistors, these little switches will also get faster every time and that they'll also draw less power as time goes on. So these are sort of corollaries. They're not part of the original Moore's law.

But really, there's an expectation that we also serve a speed aspect and we serve a power aspect as we continue to evolve the technology. So that's sort of background for your listeners. Now, I would say I can think about maybe three ways in which this has revolutionized the world. Two that are kind of very personal firsthand to me. The third kind of serves as one of my missions in my career. One of the first ones I'll go back to, I almost cried. I was a teenager.

And I would get, for reasons I don't know and can't recall, invited into these air-conditioned rooms with giant computers in them with maybe one or two people at a terminal doing something. The room was always dark. It was always freezing. I was always told, whatever you do, don't touch anything. This was what the computing world was like in the 70s and early 80s. And you fast forward 50 years later,

all of us have that exact computer in our pocket. Uh, and some of us have it on our wrist right now. You know, this is, this is the world we live in that, that technology that this, that was, you know, basically, you know, the, the, the modern day, uh, version of, of the church. It was really kind of a hallowed hall. Now everyone has, that's been democratized. Everyone can do what those very few people could do 50 years ago. Um,

Another example I'll say is when I was in college, a freshman in college, on a Saturday morning at my college, you could walk down the quad and you would see this giant truck moving at about one mile an hour. It looked like a, I mean, almost like I would call it like what a cyber truck looks like today, but it was a big truck. It was really ugly and

It would go about one mile an hour for about 10 feet. It would stop for about two minutes and then it would roll lumber along for another 10 feet. And one day I decided to ask what was going on. And they said, this is our research on self-driving cars.

And, you know, we take photographs and videos, then we process them, you know, and the grad students who were inside showed me inside. It was full of computers, full of power packs. It would go again, take pictures, drive itself about 10 feet, stop, take, you know, stock of what it had collected and take the next 10 feet. And as a freshman in college, I, you know, I looked at this thing. This is idiotic. This thing is never going to work. You know, but look at what's happened in that time frame of, you know, in the order of,

40 years, 10,000x improvements due to Moore's law in the number of transistors that fit on a chip, the power that each of those transistors draws, and the speed of each of those transistors. And that, you know, I won't shortchange all the other work that's gone around it, but that is what has enabled that monstrous box of computers and power to fit now into something the size of a glove compartment, operate at speeds where the car doesn't need to go one mile an hour and stop

And now you have very real self-driving cars built on fundamentally the same technology, but enabled by this five orders of magnitude driven by Moore's law. Third example I'll give you, this isn't as much my personal experience as something that, again, shapes my mission is,

We have really enabled a lot of the 8 billion people on the planet to improve their lives in ways that they never could, I would say, for almost all of human history. You're now in a situation where people across most of the planet can communicate with one another relatively easily, both audio and video. They can learn whatever they want to learn relatively easily. Most everything they would like to learn to improve their lives is available for free.

This has been true for some time in our country, in the United States, and in the developed world. But I think it's beginning to be broadly accessible to the rest of the 8 billion people on the planet. And I mean, I think to me, technology can be used for nefarious things that we know about, but I rely on the fact that by and large, it's a force for good. And I think Wurzla has really enabled that for the public.

Something I've always wondered with Moore's law is if it is an organic process or more of a kind of self-fulfilling prophecy, like engineers wouldn't be happy to release a new chipset unless it has doubled. Yeah, I think that's a great observation. Because first of all, it's not a law in any sense of the word we would think of. It's not a law of nature. It's not a law of physics. It's not a law passed by any government. It's merely an observation. Yeah.

And what's happened is the whole industry has taken that observation very seriously as marching orders.

meaning by definition, we will deliver 2x the number of transistors on a chip in two years. That will set our goals for the near term. It will set our sights on what the long-term goal has to look like. Intel, as I would say, we feel this as the stewards of Moore's Law, we've taken that quite to heart, but the whole industry has taken that to heart. So it drives the semiconductor equipment industry. It drives the EDA industry. It drives the Fabless industry. It drives all of us

in the same direction. And so in a sense, it's organic in that it is based on the collective execution of many, many parts that are loosely interacting with one another. But in a sense, it's a mission statement for the whole industry. Yeah, so a bit of both. A bit of both. A bit of both. But again, I think, you know, I remain very optimistic about Moore's Law for the foreseeable future.

Okay. Which brings us nicely to the next question. So many have predicted the death of Moore's law, yet companies like Intel continue to push the boundaries. What breakthroughs or strategies are keeping it alive? And how do you see this impacting emerging technologies, especially things like AI, artificial intelligence? Yeah, great question. So yeah, you're absolutely right. The death of Moore's law has been around almost as long as Moore's law has been around.

Gordon Moore himself, by the way, predicted the death of Moore's Law three times in a row. First, he said it would die in 2005. That came and went. Then he said it would die in 2015.

Uh, and then soon before he passed away, he said, 2025 is going to be the death of Moore's law. So we'll see what happens next year. Uh, but you know, I am optimistic. Um, uh, you know, it's absolutely true that, that it's been predicted and it's been wrong over and over again. And part of the reason it's wrong is the breakthroughs that we've had to do to keep it alive. I would say, um,

In the early days of Moore's law, those transistor 2X every two years were delivered by what was known as Dennard scaling. Very elegant. I'll maybe mention this. Bob Dennard was a researcher in the industry. He did some fabulous and very innovative work for all of us that said, if you do this to your transistor, it will get twice as small, it will get this much faster, and it will use this much less energy. So he kind of wrote

a playbook for us. By the way, it's also a very short technical paper from the 70s. It's a little bit of a harder read because it is a very deep engineering paper, but it lays out the math beautifully and elegantly that says, do these things and you will deliver Moore's law. And that was our playbook for quite a while, what we called the era of Denard scaling.

And that kind of ran out of steam in the late 90s, early 2000s. That path to delivering Moore's law was kind of out of gas. And then we entered what we call in our industry kind of the materials era. We innovate new materials in lieu of the scaling that Bob Dennard laid out. And new materials would be things like high K and metal gate as examples, silicon germanium, strained silicon as examples. So we'd have to go deeper into the periodic table

and find elements that we weren't used to using, find ways to use them. But ultimately what we were still trying to deliver is smaller transistors, lower power, faster transistors. Those three things didn't change. Those were the kind of pillars we were after. And then that, I would say, lasted a decade for us where we kind of mined the periodic table pretty well. You'll find us using nickel and cobalt and

you know, lanthanum and, and, you know, a lot of strange materials. Most people don't think of on the periodic. Then we kind of entered an architectural era where we would, we went from a planar transistor to a FinFET transistor. Now we're gone from a FinFET transistor to a gate all around transistor. And so we kind of continue to deliver on those exact three pillars with architectural innovation, where it's the architectural thing of it as changing the shape of things, uh,

You know, we're no longer making a transistor on a flat surface. We're making it three-dimensional. And these are the breakthroughs we've had to employ as an industry to kind of keep things alive. The emerging breakthrough, I would say, coming to us is the advancement in packaging.

I see that as one of the next emerging breakthroughs. It's been happening, so I won't say it's starting today. It's been happening, but I think that's going to carry us forward in the coming decade for sure. As far as kind of strategies, I guess there's one I am well aware of because we're kind of religious about it. We have a very long pipeline starting from early research all the way to when this is going to be on shelves in stores that consumers can go buy.

And I think we are unique at Intel in that regard. Our research to development to manufacturing pipeline, it's very long and it's very well integrated. So that research, that development, that manufacturing, to my knowledge, Intel's the only company that does it all under one roof. We do it on one site. We do it on one campus with one group of people. They do early research and that research can last a decade. We are

Today, working on ideas that the consumer may not see for 10 to 15 years. And then we have the same under the same, you know, one roof, the same clean room where we do the work. We have the development going on and then we have the initial production going on. So in that sense, I think we're unique and it really enables us to shorten the timeline between an idea and research. And how do you get it all the way to the hands of the end consumer? Yeah.

I've always in my head kind of the analogy, the most law sort of like the oil industry. So you tap all the easy fields first. And then as technology has to get more advanced, you've got to go be more exploratory. And like you mentioned, going deeper into the periodic table and then playing with the architecture is what would be the next big quantum leap.

Yeah. And I see, again, one of the emerging areas, which we, like I said, maybe I should say, we've been doing it for a little while. I think it's going to blossom, is the use of advanced packaging to forward Moore's law. And so here I have to maybe say, we maybe have to take Moore's law in spirit as opposed to the exact literal way Gordon Moore wrote it down. Because once we talk about, we're no longer necessarily talking about transistors on a single chip.

We might be talking about total transistors in a package. You know, now we're talking about that package having multiple chips on it and integrated in three dimensions, you know, X, Y, and Z. So, you know, that's a little bit of a interpretation I'm taking there to say, at the end of the day, though, what we want to deliver to the end user is those more transistors that

architects and designers can use to create functionality, each of those being done with lower power and higher speeds, those three pillars don't really change for us. But the advanced packaging era, I think, is upon us. We have begun that journey through going from very basic packaging, I'd say, for a lot of the industry, to what we call 2.5D packaging to now 3D packaging.

where chips, and we refer to them now as chiplets, because you can now partition individual functionality onto a relatively small piece of chip, silicon or otherwise. And then you can put them together efficiently through packaging. So you can use the best semiconductor technology for a certain function,

make a small chiplet that does that function, and then very efficiently put that together in an advanced package. And I think that is going to kind of be one of the ways we continue to deliver the promise that we have to deliver. For us, it is packaging technology called things like Foveros and Foveros Direct, direct copper to copper hybrid bonding,

But I will say, as an industry, we're just in the early innings here. The technology is emerging. I would say we have a long way to go to really be smart about how to use it in the best possible way. Because the idea of taking chips and putting them together in a package, that's not a new concept. We've been doing that forever. The devil's in the details here. And one example of the details is, well, you lose power when you transfer information from chip to chip.

that takes power that you know that is essentially a source of power that you have to deliver

And so there are inefficiencies that come from doing packages this way of multiple chips and chiplets. The real secret sauce is how do you partition the functionality into the right number of chiplets? How do you put those chiplets back together in a package in the right efficient manner? Now we have the ability to turn chips upside down, so you can bond chips face-to-face, you can bond chips face-to-back, you can have wires from the bottom of the chip, you can have wires for the top of a chip.

We're like a kid with a box of Legos right now. You can do anything you want. And so the real trick is how do you do exactly the right thing with all of this capability? Yes, of course. Moving on then. Some are predicted that we'll be able to achieve 1 trillion transistors on a chip by about 2030. What challenges and opportunities does this milestone present for the industry?

Yeah, I think we have a very good chance of getting there as an industry. Again, maybe I would say we think broadly, we say that I do expect the number of transistors on a chip to continue to improve. Just kind of the classical Moore's law vector. I do expect that to continue to improve. But as I said, I think a lot of it will come from how you put multiple chips together on a package.

That's how you deliver the real value. Most people don't care about the number of transistors on any chip that they own, whether it's their cell phone or their laptop. Couldn't care less, really. And I appreciate that. What they care about is what are the vectors they're getting out of it. It's battery life. It's to some extent power consumption. It's speed. And functionality. These are the things they care about. So I think that that's going to continue, that ability to deliver more and more. The opportunities, I think, is...

Look, today, you know, while we are having this conversation, some freshman at some college is looking at something and saying that is never going to work. And I expect that exact thing is the kind of thing that is going to work by the work we are doing right now.

You know, I'll say one of the, one of the things I remember is somebody told me this once is a number of inventions were predicted by Star Trek in the 1960s. You know, your flip phone was predicted by Star Trek doors that opened by themselves was predicted by Star Trek tricorders. You know, you, you,

computers you could talk to. These were all predictions from the 60s from a show that was just making it up. And they happened to be spot on with the number of innovations that we have in front of us today. We still don't have teleporters.

So maybe that's one of the things that will be enabled by this technology that's coming next. We still don't have holodecks, to my knowledge. Maybe that's going to be one of the... We definitely have AR VR. So I can imagine a holodeck might be an actually achievable goal with advancements in technology that we're working on. So, I mean, I'm being a little glib here, but the point is,

There are pretty amazing inventions that require us to continue to deliver the vectors of Moore's law to make them kind of globally available reality. And those are the kinds of things, whether you're a freshman in college saying that'll never work, or you're seeing it on a TV show thinking that's amazing, that can't ever happen. These are the things that are going to happen in a decade.

Cool. Can't wait forward. I look forward to Holodeck. Yeah. So on the subject of packaging, so innovations like Selective Layer Transfer, SLT, can help improve upon existing packaging technologies and revolutionize chip assembly.

Can you explain how SLT works and its implications, again, for AI and other high-performance applications, if possible? Yeah, absolutely. And let me put it in context first. So every December in San Francisco, we have the main technical conference for our industry, the main semiconductor conference. Because we're engineers, we don't have cool names like Davos or anything like that. It's called the IEEE Electron Devices Meeting.

or IEDM. None of that rolls off the tongue, unfortunately, but that's what it is. And it's the place where every year we all get together. We talk about our best ideas. We show the work that we've done. This past December, well, again, this couple of weeks ago, we had seven papers, which is quite a lot for one conference. Intel had seven papers at the conference.

where we unveiled some of the breakthroughs we've been doing in research. One of them was selective layer transfer that we exposed to the world for the first time. The idea is, let me say, there's really two ways to do some of this advanced packaging. I'm going to oversimplify. Bear with me on that. But one is called wafer-to-wafer bonding. You take two wafers, you stick them together,

And then you kind of cut them apart and put them into packages. The other is called chip to wafer. You take one wafer, you cut it up into chips, and then you take each chip and put it on the other wafer, one at a time. And they each have pros and cons. Wafer to wafer bonding is very fast because you're basically doing it all at once. It's a highly parallel task. But those two die, both those wafers have to have die that are the same size. Otherwise you'll have mismatches as you stick them together. Okay.

Chipped to wafer has other advantages. It can enable sort of more precision in some ways, but it's much slower. Selective layer transfer is our way of having our cake and eating it too. It really brings together the best of both ideas, chipped to wafer and wafer to wafer at the same time.

By allowing you to put two wafers together, but then not making both wafers completely stick together, but just choosing part of one wafer to stick to the other wafer.

And again, I'm using sort of non-technical terms and simplifying. I hope your audience will bear with me on that. But that's the gist. The long and short of it is I walked out of that talk. I was sitting in the audience with an executive from another semiconductor company. And as we were walking out, he said to me, I have just seen the future.

which makes me feel pretty good because that is what we're going for, obviously. But that's quite a nice validation from one of my colleagues in the industry who had no reason to say that other than, you know, I think he really did believe it. Yeah. Well, yeah, definitely. Definitely a good thing to hear from your competitors. Absolutely. Yeah, yeah.

So on that subject then, how does Intel's approach to advanced packaging compare to your competitors? And why is it critical to maintain leadership in semiconductor technology? Yeah, I'd say in one key way that I'm aware of, and I don't know everything our competitors do, but I would say in one key way, our pipeline for research to development and manufacturing is very holistic and monolithic.

Again, we do it all under one organization. We do it all under one roof. The ideas can germinate in the research setting for, again, maybe as much as a decade sometimes. If you look at thin-fet transistors or gate all around or high-key metal gate, I can say that we were researching those for a decade before we really moved to the next phase. And sometimes it takes that long to work out all of the difficult problems.

But then we have a very efficient transfer of research to development because we all work together. You know, I will take stuff we have in research and those people who have worked on it in research will go join the development team for one or two years and stick with it. And they are the experts on this technology. And then we've also got the same paradigm on the other end, the development team as they're finishing development, we'll stick with the manufacturing for quite some time. You know, it's the,

definitely enabled by the fact that we do it all in the same roof. So you don't have to drive to a different office. You don't have to relocate to another city or state or country to do that transfer. You're literally just going to do the, you know, you're coming into the same building and you're just working on the development phase instead of the research phase or the manufacturing of products instead of the development. Uh, I think that's a very powerful paradigm. Uh, and, and again, to my knowledge, kind of one that's unique to us, um,

And I think it is critical for this maintaining the leadership. You alluded to it early on, which is the parallel of the oil industry where you start with the easy stuff. I will tell you, to be fair to all of us, that it never felt easy at the time. So even when we were working on this in the early 90s,

At the time, it felt pretty hard. Now we can all look back and say, well, that was a much simpler time, quite frankly, that just followed Denard scaling and things kind of worked themselves out. And now we are doing a lot of things to deliver the value. But the key to doing a lot of those things is to start that research early and have a very seamless pipeline all the way to when it hits the shelves for customers. Yeah, so presumably you have several technologies available

developing in parallel. You don't sort of start, finish 10 years and start a new project, obviously. You've got, it's kind of staggered. The idea is to still roughly have a new technology available every couple of years. So, you know, given if the full pipeline we're talking about is, you know, 15 years for some things, then yeah, you've got a lot going on in the parallel under the same roof. Absolutely. And presumably...

Sorry, go. Sorry, Simon. Yeah, I think the other thing that I would highlight is the importance of working very well across the whole industry. It is an all play. The days are long gone, I think, when anyone can just go it alone.

and get it done. We work very closely with our equipment suppliers, with our material suppliers. We work very closely with our partners in design and in the EDA space to make sure that our interventions are compatible with how they can use them and how they want to use them and how the tools are going to be developed to support that. We don't want to create a brilliant idea that nobody can use.

So I think that's critical as well. Using, using, uh, ecosystem partners like consortia, you know, we work very closely with, um, with, uh, universities worldwide. We work very closely with consortia like, like IMEC in, in Belgium, um, to do a lot of the early pre-competitive work, the research work, uh, and, and occasionally development work that we have to do to kind of bring the whole ecosystem along. Um,

Again, not only do you not want to have a brilliant idea that you can't use, you don't want a brilliant idea that you can't manufacture because the right equipment wasn't available, for example. So as you come up with the ideas, we want to make sure that, hey, maybe this idea is going to require a new piece of fab equipment.

We better start early engaging those fab equipment companies, figure out, you know, can they make the equipment so that ultimately our vision and maybe again, this is where we're really kind of unique is even from the early research phase. We are thinking, how can we make a trillion manufacturable ones of these? We're not thinking, how can I do research and write a paper and call it a day?

Every single idea that we have coming out of my group, we have done the work ahead of time to say, is there some fundamental reason why we couldn't make millions of chips with this idea? And if there is, we better go work on it. Maybe this idea requires really exquisite process control.

So now we got to go work with all of the right parties to make sure we can get that exquisite process control to make this a reality. What we don't want to do is have the idea and say, hey, the process control is somebody else's problem or the cost is somebody else's problem. You know, yeah, this is a great idea. It's going to be expensive. You figure it out.

uh we want to we want to make sure we we're thinking all the way down to the end game from the very beginning oh that's crazy must be frustrating at times then have these kind of potential roadblocks in the way like you've got to invent calculus like what's the yeah i'm gonna say exciting yeah you you know i won't deny it's occasionally frustrating but but we frame it as a problem we have to go solve but you're absolutely right in the same way if we want to

If we want this to work, we have to do the equivalent of inventing calculus first and then using it. I don't envy you. On that subject then, so gate all around, you mentioned earlier, scaling and technology is like, is it ribbon FET? Or ribbon FET. Ribbon FET is fine. Represent major advancements in transistor design. What makes these innovations so critical for the future of Moore's law?

This is the next evolution in the transistor. And in fact, some would say it's probably the ultimate evolution in the transistor. If you could take all your knowledge of transistor physics and say, here's the cartoon version of a transistor I would like to make, it would look like ribbon fed. It would look like the channel that conducts current is completely surrounded by the gate material that turns that channel on and off. So a transistor is a switch. It's on or it's off.

uh and you would think uh you know again in a very idealistic world you'd say if i could make a channel and i could control whether it's on or off completely that would be the best switch and completely in this case means you put the gate material all around it so i would say some people would say that this is kind of the gate all around ribbon fed architecture is the ultimate evolution uh of our idea you know and and

In this sense, this is how we're now going to deliver the next iteration of Moore's law. We're going to use this with our ability to make transistors smaller,

and make them draw less power and make them faster. Breakthroughs are needed. Here again, we announced a breakthrough at the same IEDM conference. We showed that you could take one of these ribbon-fed transistors and scale the size of it by getting the gate. The gate is what controls the channel being on and off, as I said. You want to make that as small as possible. But the smaller you make it, the less energy

it is able to turn the transistor on and off.

So it basically loses its ability to control the flow of electrons in that channel. We demonstrated that you can get that gate length down to a very small number of six nanometers. Nobody has ever shown that before. We showed that you can make a very good, healthy transistor down at that gate length. Gate length scaling is one of these fundamentals that came back from Dennard's law. Dennard's scaling said, hey, you've got to scale the gate length if you want to make the transistor smaller.

I'd say a number of people in the industry were saying that's not possible. We're kind of stuck where we are. And so we showed a breakthrough paper that says we're not stuck, actually. We can get to six nanometers, which is a breakthrough. From what you're describing, I don't know why I popped it, but it sounded like

almost organic in structure, like a cell or a neuron in a brain or something. Yeah, these things begin, honestly, they begin to resemble that when you stare at enough pictures. They lose their, well, in the sense that I would say they lose their rectangularness. They become a little blobby as you continue to push the envelope. Your sharp edges that you drew in PowerPoint don't appear on the physical structure that you're building.

So yeah, I would agree with you. They kind of begin to take on an organic shape to them. Crazy. Anyway. So, yeah.

Intel Foundry has made strides in 2D NMOS and PMOS transistors. What are the key benefits of these developments, and how will they shape the next generation of chip performance? Yeah, thank you. So NMOS and PMOS are the two different types of transistors we use as switches we've used forever. Collectively, people refer to them as CMOS, Complementary Metal Oxide Semiconductors, and that is NMOS plus PMOS.

Um, and that's been again, our mainstream for since the late eighties, I think, um, that what we call 2d is now taking that channel that I just described to you, which is today made out of Silicon, you know, kind of think of it as like a

a little block of silicon in your brain, turning it into an extremely thin material that we now call a two dimensional material. So, so the thing of it is we took this block of silicon and we turn it into like a little thin sheet of aluminum foil, uh,

of a material that's not aluminum foil, but basically an infinitely thin sheet, what we call a 2D material. This is what we think is the next evolution of gate all-around transistors. We take that ideal structure. You would like to scale that channel. According to Bob Denard's scaling, you'd like to make that channel smaller and smaller, thinner and thinner. The problem is as you make it thinner, silicon kind of runs out of its ability to carry those electrons.

So we are running into sort of fundamental limits that says that block can't get any smaller or it won't conduct current anymore. Well, you switch the material from silicon to, and there's many other materials that we use and are exploring, but now you can take materials that are basically just a couple of monolayers of atoms thick, but have tremendous current carrying capabilities.

And here again, we showed for the first time that you can make a gate-all-around ribbon-fet transistor with these 2D materials with very good transistor characteristics. And we kind of showed, here's what we have, here's what the rest of industry has been reporting.

And these are kind of record results and sort of show a path to how do you continue this transistor architecture. We're now going to make a materials change where you take the silicon channel out of the middle of it, replace this with this ultra-thin 2D material. You've got to pick the material right. You've got to engineer things correctly. But when you do all of that stuff, you get an even better transistor out of it. Carbon nanotubes have any future there?

Not in this context, at least that's not the way we're seeing it. Carbon nanotubes is an area of high interest. It's running, I'm going to say my view is it's running into some fundamental problems. What we are looking at is the material systems we generally as an industry more interested in, and this is a wonky term, is transition metal dicalcogenides.

things like moly disulfide as an example that we talk about, and then materials we don't really talk about too much. These are kind of the more interesting directions. Carbon nanotubes was a material of interest and it remains a material of interest. I think my personal opinion is it's probably going to run into some fundamental issues. And I think if I just dig into the reason why, carbon is a group four element just like silicon. It's in the same column of the periodic table.

it suffers from a lot of the same problems as silicon in terms of its ability to conduct as you try to make it thinner and thinner. So, so we think that material system is probably not going to get us far enough. We got to look at some of these, what would say more exotic materials to get us there. Okay. Fair enough. Yeah.

So improving interconnect performance is just as important as transistor advancements. Can you elaborate on how substantive ruthenium and air gap technology can change the game? Sure thing. And again, I'll start with a little context. For everyone who's using any kind of electronics today, almost everyone, almost every kind of electronics, those transistors have to be connected together to do anything. So you can have a

Trillion transistors, if they're not connected, they're not going to do anything. And those connections are almost exclusively done with copper. So there's, in every chip you buy, there's miles of copper wire, very, very small copper wire, but essentially miles of copper wire, and they're connecting everything together. Copper has been with us since the mid-90s. You know, before copper, we used to use aluminum for the same thing.

And copper came around in the mid-90s as kind of the next breakthrough in interconnect technology. And it has been great for about 30 years. But as you make those transistors smaller, you've got to make those wires smaller. And as copper wires get smaller, especially kind of the dimensions they're getting to right now, they really become very resistive.

So meaning you're pushing current through the transistor and the current really can't get fast enough through the wire to get to the next transistor. So the wire is becoming your limiter and copper is really hitting a place where at these dimensions, it really doesn't have the same ability to conduct. So at these dimensions, other materials can conduct better than copper.

And one of those materials is ruthenium. And again, this is back to finding all these elements in the periodic table most people have never heard of, don't come across in their everyday lives, but they serve an important purpose for us. And I think ruthenium is one of those materials I can predict down the road some years from now, everyone's going to have a lot of ruthenium in their pocket and they're not even going to know it because it does look like a very exciting next step after copper.

So we did present a paper on ruthenium interconnects at the conference and talked about, you know, the advantages ruthenium offers over copper at these smaller dimensions.

One of the things we added to that was this notion of air gaps. So now between these ruthenium lines, the material that you have between them is very important because these lines essentially have some crosstalk to them. As you put any two wires close together, you get a little crosstalk. If anyone's coiled up a bunch of speaker wire and heard a hiss or a hum on their speaker,

This is the kind of crosstalk. That's an inductive coupling that you hear in the speaker. There's different types of crosstalk between wires. But what the material between the wires can create is a capacitive crosstalk. So these wires are kind of capacitively coupled.

They're talking, quote unquote, talking to each other, sharing information that we don't want between those wires. The air gap is one way to reduce that. So essentially we're putting, you know, think of it as air between the two wires and air has a very good dielectric constant. And that reduces the coupling between the wires. So it enables these wires to transmit information faster because that's not going to become then contaminated wire to wire.

So it's a breakthrough. I think a lot of the industry took notice and said, hey, this could be for all of us the next evolution of the wire on the chip. How rare is ruthenium?

It's plentiful enough. Put it that way. We do a lot of homework on these things up front. Like I said, when we begin researching something, we always check the, hey, is it cost effective? And B, can we make however many billions of chips we expect to make in the future? Yeah.

Early on, we do the paperwork to make sure that we're going to have an adequate supply of this. Sometimes it is a material that's very rare. Sometimes it's a material that we can anticipate that geopolitically we won't have access to. And so we're mindful of those two considerations, the rarity of it, the cost of it, the availability of it.

literally a decade before we're going to need that stuff. And so, you know, in the case of ruthenium, trust me, we have made sure we're going to have more than enough for our purposes. Okay. For materials investors, wink, wink, nod, nod. Well, we make sure. I mean, look, neon is an example. A large fraction of the world's neon comes from the Ukraine. And so we have to do a lot of work. And neon is very important in lithography in the semiconductor industry. We have to do a lot of work to make sure it's available.

Things like hafnium, things like cobalt are materials that come from countries that we have geopolitical concerns about, and we have to make sure that we can secure the supply chain. So,

It's a very real consideration in the industry to make sure that as we go away from the most abundant elements on the periodic table, we're really sure we know where they're coming from and that we can get them. Yeah, absolutely. Absolutely. Okay. As we approach the physical limits of our current technologies, what do you see as the next major innovation needed to achieve ultra low power consumption and meet the demands of AI? Yeah, that's great.

a great question and great thing to tee up. I would tell you one of the things that keeps me up at night, keeps us up at night is throughout all of the innovations created by Moore's law, one trend is troubling, and that is the consumption of power by computation and communication worldwide is growing at a much faster rate than the supply of power in the world.

Roughly speaking, and you can see this whether you talk about the US or Europe or worldwide, the trends are similar. The total computation plus communication power is growing at about a 25% compound annual growth rate. Driven a lot by AI today, but that growth rate has been fairly steady and this predates AI.

but it's been growing at this 25% year-over-year compound annual growth rate. At the same time, the supply of energy, supply of power, has been growing worldwide at about a 3% compound annual growth rate.

Those two lines are going to intersect at some point. And what you're starting to see now is beginning to be some side effects. You talk about companies building data centers and then talking about putting a nuclear power plant right next to their data center. Or you talk about companies that say, we're going to build our entire data center completely underwater because it'll be cheaper to cool it. And innovative ideas that maybe work, maybe don't work. So I think we're

We're running into a, we're heading for that wall in that sense that, you know, if you just draw those lines today, they intersect about 2040. So not that far away. Wow. Okay. You know, in 2040, if we, if we don't do anything, then all of our worldwide power supply will be going into computing and communication. No, no power left to eat anything, no power left to fly anywhere or drive a car or any of that stuff.

Obviously, that's not going to happen. We're not going to sacrifice other things. What's going to happen is something different. From my point of view, I know plenty of people are working on the supply side of that equation. How do we supply more power?

I think one of the calls to action that I believe in, and I've been advocating in the industries, we need to work on the demand side. How do we deliver the same computation and demand less energy for doing it? And there's some natural Moore's law aspect to it. Again, each of these transistors is expected to use less energy.

and has been for 60 years now. So there has been ongoing, but even with that ongoing improvement in energy consumption per switch, we're on this upward trend because the demand for that computation is outpacing. So one of the things we're invested in and very interested in, and one of my colleagues at the conference gave a nice talk about, is the idea of making ultra-low power switches. So a dramatic, really, step function decrease

in the amount of power that one of these switches would take. These switches would be based on really an entirely new type of physics from a traditional transistor.

And frankly, they would require, in my opinion, a rethinking of the computation stack top to bottom. You'd have to rethink, how do you build chips with this switch? How do you build a chip architecture with this design methodology? What's the software that sits on top of this? The whole paradigm would be completely different, I would say, from the classical von Neumann computing paradigm that we have today. It's an exciting area. It has huge potential.

It's also a daunting amount of work for the industry. You know, and this is where I think one of the roles we serve at Intel is to kind of set that direction for the whole industry.

This is coming. Here's all the things it's going to take to make it successful. And you can't do one thing in a vacuum. We can't, for instance, just invent this new switch and stick it into an existing chip design. That probably isn't going to work. You have to reimagine the chip design. It'd be like taking a house made out of brick and then swapping all the brick for blocks of ice.

you're not going to make a different kind of house that way, right? It's really, you know, you wouldn't build an igloo the way you would build a brick house. So it's the same way. We're going to have a different type of switch in the future based on a different type of physics, far more energy efficient, but operating really in a different way. And that's going to require how do you then build everything above that till you get to a user experience has to be reimagined as well. Yeah, yeah, absolutely. Yeah.

I think I saw with some AI proposals to make them faster and less power consumptive is to have

kind of algorithms hard printed, hard coded onto a transistor or something. Would that be potentially a way of reducing power consumption? Absolutely. And I would say the industry has been somewhat going in that direction for some time now in the sense that we're more and more taking algorithms that were implemented in software. And when they're sufficiently, I think, robustified,

take that algorithm and implement it in hardware. Loosely, the term for this is ASIC, Application Specific Integrated Circuit. Yeah, that's the term. And more and more, I would say, products are coming out looking like these ASICs. So instead of maybe doing AI machine learning on a truly, fully general purpose computer, you get more and more purpose-built

And essentially what you're doing is you're using exactly the number of transistors and exactly the type of transistors and wiring up in exactly the right way to achieve the one function you're trying to achieve. And that's how you can do all of that function without any wasted power. And that trade-off is now you've got something that's good for one thing or a couple of things, but it's not good for everything, right? The power of a CPU is that it's highly general and you can build a lot of things out of it.

As you get more and more specified, it does one thing really well and efficiently, but it doesn't do anything else. The world, I think, is going to exist with both in the end. Yeah, that would lead more to a modular computer system that you can swap out. Basically like your old Game Boy with your game cartridge. Yeah, and here's where I'm back to chiplets. Now you can have some of this functionality truly hard-coded in a specific chiplet. And also, I would say,

optimized with the type of semiconductor process that that chiplet needs. And then you can have other functions in the package built out of different chiplets, and part of that can be a CPU on a most advanced leading-edge technology. Some of it can be DRAM or NAND or SRAM memory from a node that's optimized for that. So now you can kind of harness...

the most energy efficient way to do different tasks. Again, the secret is putting that all together in a package without squandering all that benefit in the package to pack it, you know, the die to die transmission loss and things like that. That's easy to do. I would say every one of these things looks great in PowerPoint. And the challenge is to make them really work well. Yeah, absolutely. Fantastic. So moving back to the resource subject for our final question.

So Intel is championing the effort to strengthen US semiconductor supply chain. What are the key benefits of regional diversification and how does it support global resilience in the industry? Yeah, I would say this. I think the need for a resilient global supply chain is almost more clear than ever right now. You know, if you just...

the geopolitical conflict going on around us and you map the threats of geopolitical conflict that we've been hearing about, you know, the steady drumbeat of those threats, the world demands a geopolitically secure supply chain so that you have materials coming all the way from top to bottom in the stack from enough parts of the world that we're not going to be impacted by any one geographic location basically supplying 90 plus percent of the

of that section of the market. That's absolutely required. I would say when I look at it from a U.S. context, the U.S. is obviously working diligently to rebuild its domestic supply chain. The CHIPS Act was one example of that. I would say Intel is knee-deep in the middle of that. I think, as I said, we consider ourselves the stewards of Moore's Law. We're kind of, in the U.S. context, the only company that does leading-edge manufacturing in the U.S.,

We do all of our semiconductor manufacturing in the US and in Europe. We don't do any semiconductor wafer manufacturing in the Far East. We do some packaging in the Far East, but we don't do any of that. And I think that the leading edge semiconductor world, which is very important, say, to AI and to the extent that AI is important to geopolitical considerations in the future, I would say that supply chain today looks very unbalanced.

And so from our point of view, bringing balance back to it, bringing more leadership away from manufacturing to the US, to the extent that it services the AI market, to the extent that the AI market is, again, critical for the future of geopolitical safety and security, I think it all comes back to the same point.

Absolutely. That's the end of my question, Sanjay. Is there anything else you'd like to add before we close up? I really enjoyed this, Chris. And I think those questions were spot on, first of all. And I appreciate you giving me the chance to wander into some wonky backstories here and there to lay out some context. Again, I think we had seven papers at IEDM. You allowed me to talk about a few of them, and I appreciate that.

Broadly, the theme, though, is we're continuing to innovate the research for the whole industry, not just for us. And in examples like IADM, we're making our results public. You know, this isn't some secret we're keeping to ourselves. We're sort of showing the whole industry this is the path forward.

Fantastic. That concludes this episode of Lexicon. Thank you all for tuning in and being our guest today. Follow our social media channels for the latest science and technology news. Also, don't forget to explore our IE shop. Goodbye for now.