Semiconductors and Chris Miller’s Chip War: The Global Battle over the Future of AI

Prepare to be enthralled by Chris Miller, the esteemed author of "Chip War," as he joins us to unravel the complexities of AI's impact on the semiconductor industry. We tackle the monumental task of rethinking computing efficiency in the face of supercomputers and data centers that gulp power like never before. Chris illuminates the path ahead, where technological advancement must reconcile with environmental imperatives, addressing the thermal hurdles that come with more powerful chips and the revolutionary changes needed to test these energy behemoths.

Venture into the semiconductor saga with us, tracing the arc from the inception of computing to today's digital omnipresence. The vital role of lithography, especially the marvel of Extreme Ultraviolet (EUV) technology, comes to the fore as we dissect Intel's early gambles and ASML's exclusive reign in this sphere. It's a tale of triumphs and trials, where the relentless pursuit of progress meets the gradual slowdown of Moore's Law, pushing the industry towards groundbreaking R&D investments and international competitiveness.

Our exploration culminates with a striking contrast between the human brain's efficiency and the voracious power appetite of modern AI models. As we peer into the future, we speculate on the promise of biological computing, like DNA-based systems, and how they might disrupt our silicon foundations. The episode concludes by considering how AI advancements, exemplified by the capabilities of large language models such as ChatGPT, could redefine the workforce and bridge the existing skills gap. Join us for this enlightening journey through the evolution and future innovations of the semiconductor industry with the insightful Chris Miller.

Thanks for tuning in to "Advantest Talks Semi"!

If you enjoyed this episode, we'd love to hear from you! Please take a moment to leave a rating on Apple Podcast. Your feedback helps us improve and reach new listeners.

Don't forget to subscribe and share with your friends. We appreciate your support!

Chapters

• 0:29 Introduction: Welcome to Advantest Talks Semi with Chris Miller
• 1:27 Chris Miller: From Academia to Author of "Chip War"
• 2:10 The Semiconductor Industry and AI's Transformative Power
• 3:11 The Historical Impact of Intel and Moore's Law on Technology
• 4:32 Future Challenges: Power Consumption and the Sustainable Supply Chain
• 5:45 The Evolution and Future of Moore's Law in Computing
• 6:45 Advanced Packaging and Its Role in the Future of Semiconductors
• 7:09 Thermal Challenges and Innovation in Semiconductor Testing
• 8:00 The Origins of Integrated Circuits and Their Evolution
• 9:24 The Inspiration Behind "Chip War" and Its Research Journey
• 11:19 Lithography Wars: The Pioneering Investments That Shaped the Industry
• 13:15 ASML's Role in Semiconductor Fabrication and EUV Technolog
• 16:51 The Global Economic Impact of AI and Semiconductor Innovation
• 17:52 - Bifurcation of Global Supply Chains and China's Role
• 20:38 Trade Restrictions with China: Effects on the Semiconductor Industry
• 23:21 The Massive Investment Required for the Future of AI
• 24:37 The Long Journey of Writing "Chip War" and Industry Insights
• 30:45 Advanced Packaging and Its Role in the Future of Semiconductors
• 35:38 Heterogeneous Integration and the Business Challenges of Chiplets
• 38:25 Exploring Biological Computers and Future Computing Paradigms
• 41:10 Workforce Development and the Unexplored Potential of Semiconductors

Transcript

Keith: 0:02

Welcome to a special episode of Advantest Talks Semi. In today's episode, we have the distinct honor of hosting a very special guest, Chris Miller, the acclaimed author of the New York Times' bestseller "Chip War: The Fight for the World's most Critical Technology." This book explores the intricate web of technology, power and politics of the semiconductor industry and also earned a spot on Barack Obama's list of favorite books 2023. Chris Miller teaches international history at the Fletcher School at Tufts University and holds the position of Gene Kirkpatrick Visiting Fellow at the American Enterprise Institute. Additionally, Chris is a director at Green Mantle, a macroeconomic and geopolitical consultancy based in New York and London. His insightful contributions frequently grace the pages of the New York Times, the Wall Street Journal, Foreign Affairs, Foreign Policy and the American Interest.

Keith: 1:00

Chris received his PhD in history from Yale University and an AB in history from Harvard University. Stay tuned as we uncover the critical importance of semiconductors in shaping our world and the intense competition to control this essential technology. Chris, welcome. I want to talk about what's going on in the semiconductor industry. You've obviously written a New York Times bestseller that got everyone's attention, so I want to talk to you about some specific pieces in the book, but also just what we might think about is coming in the near future.

Chris: 1:39

Well, thank you for the opportunity to join the podcast.

Keith: 1:43

Okay. So, Chris, I want to start with the recent Intel design services conference that completed, I think, a week or two weeks ago, and Pat Gelsinger, the CEO, he had this quote I want to read to you. He said: "AI is profoundly transforming the world and how we think about technology and silicon that powers it. This is creating an unprecedented opportunity for the world's most innovative chip designers and for Intel Foundry, the world's first systems foundry for the AI era." I started with that quote because one it directly relates to your book, but also Stuart Pann came up on stage later and he quoted you as saying, "Intel is the most important company of the last 50 years." So just wanted to get your thoughts on where did that quote come from and how do you think about Intel as being one of the most important companies?

Chris: 2:44

Well, I think, if you look historically, the key feature that has transformed everything we know over the past 50 years whether it's economy, our society, our technology everything has been Moore's law, the tremendous race forward computing power that the semiconductor industry has delivered. Especially in the early days, the industry Intel was really at the center of that, driving it forward. And, of course, Moore's law was named after Gordon Moore, who was one of the co-founders of Intel. That wasn't Intel alone, of course. There were many other companies that were providing the designs or the tools or the chemicals or everything else that's necessary to make better and better chips. But I think the basic dynamic that Moore's law and better computing have changed everything is something that we actually take for granted.

Keith: 3:35

Yeah. So one of the interesting things that they talked about was where these three tiers the world-class foundry, which depends on performance, power, area and cost, but then two, the resilient, sustainable supply chain, which I'll come back to in a second, and then three, the systems of chips. Let's talk a little bit about the resilient, sustainable supply chain. So let me give you some context here.

Keith: 4:05

So they talk about today's latest supercomputers. They have upwards of 100,000 CPUs and they're the size of aircrafts, 747s, and they use the equivalent power of the entire state of California. So, if I think someone did a calculation and if you add up all the data centers and look at how much power is being used, it's the equivalent of around four countries worth, so the Netherlands or Sweden, or four Californias, and that's in today's power consumption. But then it goes on to talk about, in the future we're going to have, let's say, a million CPUs or maybe even 10 million CPUs. So the point I want to make here and draw your attention to and comments on is we're using currently four countries worth of electricity, but we already have this vision of being 10 times as large, so 40 countries worth of electricity. So how do you feel, or how do you think about the sustainability with that sort of backdrop?

Chris: 5:18

Well, I think, if you go back to Moore's Law, where we started the conversation, most people know Moore's Law as the doubling of computing power every couple of years.

Chris: 5:28

What most people don't know or forget is that Gordon Moore also predicted a declining average cost per transistor, which implied a declining average cost per unit of computing, and that was delivered for most of the history of Moore's Law, and I think today it's getting harder than ever, because it does get more expensive to shrink transistors further and to provide all the capabilities around them.

Chris: 5:54

But I think it is a challenge that lies in front of the entire industry, and it's a challenge both in dollar terms, but also in terms of electricity consumption, and if we don't find ways to make the economics work, then we're not going to have access to the computing we imagine, and so that's why I think we have to, and companies already are focusing a lot of their R&D right now on precisely this question how do you squeeze as much compute per unit of power you're using? Because certainly customers in the data center business are absolutely demanding it, and for them, as you mentioned, one of their key sources of cost is energy, is electricity, and so that's why they want to drive down the power consumption to the extent possible, and that's why there's so much research right now going into how can you do that in the most efficient way.

Keith: 6:42

Yeah, I know, for Advantest, one of the key challenges that we have in the test segment is thermal, is exactly as you stated, with power, and that these server chips are already at one kilowatt and they have two and three kilowatts on the roadmap over the next several years, and figuring out how to test that before it's in its final form is extremely challenging for the semiconductor industry. Back in the 50s and 60s there really wasn't even the term semiconductors or microchips or integrated circuits, and you spent some time in the early chapters of the book taking the audience through that, so could you just give you a few points from those early chapters?

Chris: 7:33

Well, the earliest computers that were created in the 1940s and 1950s were very large machines. They required extraordinary amounts of power to use them and created vast amounts of heat, and all of those were dynamics that made it difficult to shrink them down in size and deploy them on different types of devices. And so to shrink computers and to make them more energy efficient, a device called the integrated circuit was first created in the late 1950s, more commonly today known as semiconductors, which was a way of making the circuitry smaller and thereby putting more computing power in each computer and doing so at lower cost, using less energy. And that basic dynamic has continued all the way up to the present day. Advances in the chip industry have always been about shrinking computing down so we can use more of it at lower cost on more types of devices, and that's why it's not just computers that have computing power embedded inside them. It's dishwashers and coffee makers, and phones and automobiles and almost everything else.

Keith: 8:41

I think we talked about this a little bit at the PDF user conference but where did the idea of this book come from and how did you go about? I think you spent eight years working on this book, so talk us through about where the idea came from and how you went about it.

Chris: 8:57

Well, I initially wanted to understand the evolution of technologies that governments placed great priority on because they were relevant for military systems and during the Cold War in the 1950s and 60s, this was things like long range missiles and rockets that went into space, satellite systems, which were relevant for military uses, and I wanted to understand well what were the policies, what were the dynamics that led certain countries to have more rapid technological development than others. I came to realize that actually, the technology that had been the most impactful wasn't rockets that flew into space or airplanes that flew faster. It was actually the ability to deploy computing into a whole range of different types of systems, which is why, if you rewind the clock to the 1950s, almost nothing had computing power inside of it, whereas today almost everything does, and that's been the key technological transformation of our time.

Keith: 9:49

Yeah, so totally agree with that. It's amazing how prolific it now is. I mean, it's in your garbage cans. So I want to jump back to a part of your book. I think the chapter was titled Lethography Wars and I think, as you said, in '92, Intel was the world's biggest chip maker and John Carruthers had just asked the then time Intel CEO, Andrew Grove, for 200 million, which it's kind of funny because 200 million doesn't sound like a lot of money now when people are talking about hundreds of billions and now trillions of dollars. But still, 200 million at that time was a lot of money, and it was a heated exchange about what are we going to invest this 200 million in if we don't know that it works. But that was the essence of lithography. So I want to hear your thoughts on how that impacted the industry and where we find ourselves now in terms of lithography capabilities.

Chris: 10:52

Well, I think one of the factors that differentiates the semiconductor industry from most others is that it spends so much on research and development. Only really the pharmaceutical and biotech industry come close. And the industry needs to spend that much because it's developing technologies that are quite literally bending our knowledge of physics and trying to do so in a way that is cost-competitive. And there's been many advances that have been required to make the current chip so we take for granted possible. But advances in lithography have been a key one of those, and in the early 1990s there was a real debate about what would be the next type of lithography that could print smaller and smaller transistors and other components on chips.

Chris: 11:37

And, as you said, some employees at Intel went to Andrew Grove and said let's bet on extreme ultraviolet lithography with a timeline of at least a decade and perhaps more before it would ever be implemented in the practical sense and with a very large price tag for something that would only be useful after the careers ended and everyone involved.

Chris: 11:57

Andrew Grove was retired, John Carruthers was retired, everyone involved was retired before the first EV scanners were ever installed and in most industries would have no difficulty knowing what the answer to that request would be, and most industries answer would be "absolutely not. We're certainly not going to spend $200 million on a far-off project like this that might never see returns. But, to the great credit of the semiconductor industry, it's regularly willing to make investments like that because they're not just seen as important, they're seen as necessary. Unless you make bets like this, technological progress will slow and companies rely on progress for their competitive edge. And so to me, that's what makes the semiconductor industry so exciting is that there's just this commitment to constant progress, even when the progress is far out in the future and when it requires extraordinary investment to actually realize.

Keith: 12:48

So, Chris, you talked about the competitiveness of lithography, and I think we would be remiss if we did not discuss ASML and its strategic importance to the global supply of semiconductor fabrication. So in the book you have a really nice quote. I'm going to try it here: "The miracle of EUV, or extreme ultraviolet, as you said, and it's a miracle that the latest semiconductors work at all, which, being in the test industry, I can tell you we also are often surprised, and when we're testing these things we also think that it's a miracle. So I do want to understand a little bit, though, like what's behind that quote, and can you walk us through ASML's rise to EUV dominance, and why is there only this one major company that has the capability to even provide EUV capabilities?

Chris: 13:45

Well we spoke about how the early investments in developing the EUV were taking place in the beginning of the 1990s. It wasn't really until the mid to late 2010s when the first lithography tools using the EUV were beginning to be installed for high-value manufacturing. It was a 25-year development cycle for these tools, which explains why only one company was willing to see that from beginning to end, and it took so long, both because it was delayed, because it was expensive, but also because the engineering involved was just mind-bogglingly complex, and this is, I think you know, ASML is one example of this. You actually see this across the entire semiconductor supply chain, but it's a, I think, a vivid example of the complexity involved. So, tools with hundreds of thousands of components and the flattest mirrors humans have ever made, one of the most powerful lasers that was never been deployed in a commercial device, and an explosion happening inside of them, as tiny balls of tin are pulverized into a plasma that measures many times hotter than the surface of the sun, and so if you were to ask someone off the street what's the likelihood a machine like this ever works, they would say it sounds pretty implausible.

Chris: 14:59

But, of course, machines like this and basically all the equipment used in semiconductor manufacturing and testing, they have to work most of the time because they're expensive tools and if they break down it's a very expensive problem to deal with, and so the miracle is that actually, it's possible to bring together so many components and software and materials all of which is among the most complex that humans have ever done and to orchestrate it all so that it functions. When you flip the ON switch, you know it's going to be on most of the time and, starting from scratch, you or I would not expect that it's actually possible to do the things the industry does on a regular basis.

Keith: 15:35

Yeah, the latest one from ASML was just delivered to Intel a few months ago. So I think the price tag on that was upwards of $360, $380 million and, like you said, it's a miracle that it works because of all the complexities that you described, but it is required for these next generation AI high performance computing chipsets. So, we talked a little bit about the power hungriness of the chips and how we really need to design innovative, let's say, techniques or new designs that can bring that power curve way down. So, I actually want to shift gears and talk about that a little bit.

Keith: 16:24

In a previous podcast I talked about the potential economic opportunity. So Pricewaterhouse Coopers is on the record saying that it's somewhere like $15.7 trillion, $16 trillion economic opportunity, which now has galvanized the entire industry and we have all of these hyperscalers now, like the Googles, the Metas, the Amazon. They're now getting into the game to make their own chips to compete with the likes of NVIDIA and AMD and Intel and others. But I just want to get your take on is the opportunity that large? And, if it is, I've also seen that China is expected to benefit the most from that opportunity and I'd like to hear your thoughts on do you think that's true and if it is true, how would you advise countries like the US and the EU to augment or to adjust their strategies?

Chris: 17:25

Well, I think we're at a moment right now with AI where sizing where the market will be in 10 years is just extraordinarily difficult. No one 24 months ago would have predicted this splash that ChatG PT made. Hardly anyone 24 months ago would predicted that NVIDIA today would be the third most valuable company in the world, and I think that speaks to just the real uncertainty about how AI will develop in the coming years, because we were so bad at predicting it over the past couple of years, and so to me, I think that's exciting. It's really fascinating to watch all the new AI products that are being tested right now, but it also makes it difficult to predict specifically. I think there's no doubt that AI will be a transformative force. There's no doubt that will acquire a whole lot of chips to both train AI systems and then to deploy them, but exactly what scale, I don't think anyone really knows for sure, but it has, as you say, driven a surge of new investment in the infrastructure the hardware infrastructure that's needed to actually run AI systems, and that's been to the great benefit of many semiconductor companies.

Chris: 18:33

I think, if you look at the geography of how this plays out, what you're going to see and what you really already are seeing is that there are two different supply chains that are beginning to form, not completely separate, but somewhat split, partially for economic reasons and partially for regulatory reasons.

Chris: 18:51

One is focused on the Chinese market, where a number of US firms are prohibited from selling cutting edge AI chips, and so you're already seeing a Chinese ecosystem that's going to be spurred on in its development because of these types of restrictions.

Chris: 19:06

And then in the rest of the world, I think you'll have the bigger multinational players competing amongst themselves for market share and for dominance of these new markets. And I think this partial bifurcation into two different AI spheres and two different semiconductor spheres is going to continue. And you see basically every government now taking steps trying to support their own ecosystem, and I think what's interesting to watch is different governments are placing bets in different parts of the supply chain. So, for example, in Korea, where the country has very strong memory chip producers, there's extraordinary focus on the role of high bandwidth memory in AI, for very understandable reasons. You don't have the same conversation in the United States, where there's one memory producer, but there's many more companies doing other things in the supply chain. The Netherlands you won't be surprised to hear that the conversation focuses a whole lot on lithography, given the role of ASML, and so you really see very differentiated strategies among different governments, even though the level of focus is quite high, no matter where you look.

Keith: 20:11

Yeah, you mentioned the term spurred on innovation in China. So, just kind of as a follow on, do you think the trade restrictions with China are having the impact that the US government desires, or is it really, as you said, spurring additional, let's say faster innovation from withinside China?

Chris: 20:35

Well, I think it's a complex question to answer. If you were to think of it from a theoretical lens, you'd say the best type of restriction is one that sets the bar just above what you know your adversary can produce. Because you know, if you set the bar below what they can produce, they're going to produce their own and replace you, and I think we're going to learn over the next couple of years whether the US government has set the bar too low or too high and whether domestic Chinese firms can really displace NVIDIA and other companies. I think that the next question you'd ask is well, regardless of where you set that bar, what's the impact on the AI ecosystem? Because ultimately, the restrictions aren't really intending to impact the chip market. That's sort of the tool they're trying to use to impact the AI market.

Chris: 21:23

And here I think it's undeniable that US firms are in the lead in the AI race, not primarily because of the restrictions, just because they were in the lead to start. And it does seem like the restrictions will cause some challenges for Chinese AI firms as they try to build bigger and better data centers, to train bigger and better AI systems. But how important these restrictions look in the next couple of years, I think it's still pretty early to draw any confident conclusions. I'm sure that Chinese AI firms would prefer to be able to buy whatever type of chip they like, but how significant they'll be I'm not sure.

Keith: 22:00

Yeah, it will be interesting to see how this plays out. So let's pivot over to open AI and something that Sam Altman said. He's been in the news lately. He was over at the United Emirates talking about trillions of dollars needed for the future vision of AI, and there's just a tremendous amount of infrastructure and there are lots of problems and challenges still to be solved. With this number, you know, five, six, seven trillion dollars, who even has the capacity or the capability to provide such funding? And the only ones that sort of come to mind are these sovereign wealth funds. So I just want to get your take on: Do you think that would be a potential avenue that those funds might take and, if so, how might that change the landscape?

Chris: 22:54

Well, I would say that $5 trillion sounds like an extraordinary amount of money.

Chris: 22:58

I've seen some estimates and the entire amount of capital expenditure in the history of the semiconductor industry has estimated around $1 trillion.

Chris: 23:07

That's all investment from the 1960s up to the present.

Chris: 23:10

I think if you look out over the next 20 years, for example, suddenly $5 trillion doesn't start to seem like a very crazy number. However, Sam Altman hasn't ever put a time horizon on the amount of time it takes it'll take to spend that money. And if you think over a decade or two, suddenly that amount of spending might actually be reasonable. If you're looking not only at the manufacture of the chips themselves but also the manufacture of the construction of data centers and then the power plants that would be needed to power those data centers, put all that infrastructure together and suddenly you're looking at some pretty expensive processes and I can give you an idea that we're going to be deploying AI to basically every aspect of our economy, every aspect of our life, or in the next couple of decades, you can start to envision some pretty substantial numbers. Not all that needs to be funded by open AI itself, but the entire computing industry. That does have a lot of investment to undertake if, in fact, AI is going to be anywhere near as pervasive as Sam Altman thinks.

Keith: 24:10

Yeah, you make an excellent point about the time horizon. So with that perspective, then it's the five trillion is not such a big number. I mean it's still a big number, but it's reasonable. Okay, let's go back a little bit to Chip War. I wanted to first thank you again for writing such an amazing book. I did hear that it took somewhere around 8 to 10 years for you to write this book. That's correct?

Chris: 24:39

It took that long to research. It was slightly shorter to write, but the hard part was to learn about such an extraordinarily complex industry.

Keith: 24:46

Yeah, I mean you've done it much faster than I have. I've been researching it for 30 years and I still don't understand it, so you already beat me. I do want to ask over the course of the research, so you obviously interviewed the heroes of the semiconductor industry and, if you could, could you share any memorable interactions or insights? What was the most interesting, most, let's say, challenging? Take your pick.

Chris: 25:16

Well, there were so many fascinating ones, some with people you described as the heroes the CEOs, for example, or the Nobel Prize winners, but I had just as much fun interviewing folks who were in the mid levels of companies that walked me through their technology in extraordinary detail or explained what exactly their tool did.

Chris: 25:33

I think my most interesting set of conversations was with the founder of TSMC, Morris Chang, who was a major character in the book, in part because he's been a major player in the chip industry since literally the earliest days in Texas, when he started work at Texas Instruments in the late 1950s, and today he's, of course, now retired from TSMC, but played a central role in its founding and in its rise to become the world's largest chip maker, and to me he represented so much about the industry's history, its origins and manufacturing.

Chris: 26:10

He was a critical role in the assembly lines at Texas Instruments all the way to the splitting of design for manufacturing with the foundary model that's so prevalent today, and then the central role that Taiwan plays in the chip industry as well is doing no small part. Thanks to him, and to me, I came away from that research thinking everyone knows who Bill Gates is and everyone knows who Steve Jobs is, but most people couldn't name anyone from the semiconductor industry, and it seems to me that Morris Chang is a name that ought to be mentioned alongside those other, much more well-known titans of technology.

Keith: 26:44

Yeah, definitely, and his name was mentioned several times during the Intel Foundry Services Conference and I think the vision that Intel announced is by 2030, they want to be the number two foundry. So to your point, he's done such an excellent job of building up TSMC that even the greatest companies in the world are trying to become number two not actually supplant TSMC, at least not just yet.

Chris: 27:14

Well, I think that that is an extraordinary compliment to what he's done.

Keith: 27:19

Yeah, extraordinary compliment, well said. Okay, then, just this is a related question. Maybe you can give another story. But as you were researching Chip War, what was an unexpected challenge or what was the most challenging aspect of researching such a complex industry and how did you overcome that? Was it just trying to get the interviews or just trying to piece it all together? I mean, what did you find the most challenging?

Chris: 27:48

You know, I think the hardest part was to understand where did the chip industry actually end? If you ask most people what's the chip industry, they'll say, well, it's people who make chips, and that's true. And then you ask, well, how do you make a chip? And you get sent down an extraordinary journey of material suppliers and tool manufacturers and testers and designers and EDA software companies, which was great fun to learn about, and I didn't even get into the people who were actually doing the mining and processing of the silicon itself. You could go all the way, all the way that far, if you wanted to, and it became clear that actually a substantial share of humanity was actually participating in the chip making process in some way or another.

Chris: 28:31

The ventilation machines inside of a fab, extraordinary capable ventilation machines. You could go on and on and on. And that's not surprising, because making chips is so hard and so complex that you actually need to tap into the expertise of hundreds and hundreds of companies across the entire supply chain in many different countries, and thousands and thousands of very bright and hardworking individuals actually make all of this come together. And so to me, that was the greatest challenge was to understand try to understand the complexity of the supply chain and I did my best to illustrate it as far as one could in 300 pages. But there's so much that I could have added in if I had more space to touch on other parts of the supply chain that were possible to bring into the story.

Keith: 29:18

So, Chris, it's interesting that you say that because we have career roles inside of Advantest, that's, their single job function is to try to make sense of the complex supply chain, and multiple people across the company in different market segments spend their entire careers constantly tracking and understanding that. So, to your point, it is extremely difficult and challenging. Let's bring that to the present. One of the things also that Intel talked about at the conference were the advanced packaging. So you mentioned the materials and how, really the material physics, which was always important, but I think today it's probably, well for sure it's much more important today than it was, I'd say, 20, 30 years ago. It may even be the most important thing in terms of competitiveness moving forward.

Keith: 30:18

So, there's these new 2.5D, 3d IC packaging and companies are already looking at how to integrate optics and silicon photonics into the packaging. Because, as we talked about earlier, the power, the thermal at some point we hit a brick wall. As the data bandwidths get higher and the performances increase, we can't just continue to add power. There are all these new advanced packaging techniques and one of the more interesting ones that Intel was talking about was glass substrates and glass packaging that will be used to integrate the photonics and the optics into the SOCs. So, just wanted to hear your perspective on 2.5 and 3D. I mean, are we going to see 4D? What's beyond 2.5 or 3D? What do you think?

Chris: 31:22

Well, I think one of the challenges of the entire industry is dealing with is that the cost declines associated with Moore's Law have slowed over the past decade or so, and that's understandable, given that the transistors in your iPhone are now basically virus sized. It's harder and harder to make them smaller while also making them cheaper, and so the entire industry is looking for other ways to gain efficiency improvements, and there are two main pathways for this right now. One is on the design side, where we've seen, I think, an explosion of new design types, of which the GPU is probably the most high profile example, but there are many others. And then the second is on new types of packaging, and, of course, advanced packaging itself refers to hundreds of different types of techniques. It's sort of a deceptive slogan because it means so many different things, but certainly I think you see almost every company exploring new ways to bring different types of semiconductors closer together and more efficiently together. Together in ways that dissipate heat more effectively, and I think that's going to be, as you say, a major driver of both technological progress but also helping on the cost and the energy efficiency side, where shrinking transistors alone is not giving us the gains that we were used to seeing over the history of Moore's law, and that's why there's so much focus on it.

Chris: 32:50

I think the interesting question about the revolutions in packaging that we're seeing is how will it impact business models? Because traditionally, packaging was something that was done by a part of the supply chain that was separate from fabrication, and now, actually, what we're seeing is that a lot of the companies that are investing the most in advanced packaging are actually people who are on the front end of wave fabrication Intel, TSMC, Samsung and others. And what's even more interesting, I think, is the extent to which some of the memory chip firms are also focusing on advanced packaging services, because they're saying we're going to provide the high del of memory, you give us your accelerator, we'll put them next to each other. And so I think we're at that point where we could see some pretty interesting changes in business models over the next decade, driven by the way packaging becomes a more important part of the value add, and that could be a big factor in shaking up how the industry is structured.

Keith: 33:44

Yeah, let's unpack that a little bit, because a couple of things there of note.

Keith: 33:48

With the advanced packaging we're really talking about a heterogeneous integration and chiplets, and that actually brings us all the way back to what we first talked about when we started this discussion, which was systems of chips. So that was Intel's third tier. So when I was testing SOCs in the late 90s and throughout the 2000s, it was all about system on a chip, where we're taking different IP blocks and putting them together to create a system on a chip, and we were talking about what sort of impact and challenges that was going to create for the test industry. But now we're talking about systems of chips and, to your point, we're taking different chips, known good die chiplets from different companies and entities and putting those together in these elaborate, complicated systems. So when you do that and something doesn't work, who pays for it? Who owns that product, or who owns the troubleshooting, or who owns the time to market, or who owns the failure? So, like you said, it creates these new business challenges that I can tell you the semiconductor industry is still struggling with today.

Chris: 35:11

But I think that speaks to the challenge of standardization.

Chris: 35:15

How do you have standard interfaces, standard processes to facilitate this type of chiplet integration? When it came to IP blocks, like you were talking about, over the past several decades you saw the EDA tool providers play a very big role in that integration process for example, and I think we're still in the exploratory stage of it was going to be the driving factor in standardizing, and will that require companies to change the way they operate to do that? I think they're probably major new markets that will be opened if we're able to standardize the chiplet process in a way that allows mixing and matching to the ultimate degree. And I think one of the most exciting dynamics is that it could bring new types of companies into the business of designing their own chips.

Chris: 35:59

Right now it's the big tech companies that have vast budgets Meta, Google and others that you mentioned, but in the future, if we're able to use chiplets to drive down the cost of designing your own system because you can if you have 20 chiplets, you can buy 19 on the shelf and just design your last one for your own use case then suddenly the cost might be low enough for other types of companies and other industries to also design their own, and so I think one of the test cases of whether the chiplet revolution really takes off is whether we get more auto companies or medical device companies or industrial companies designing their own systems, and to me that would be a pretty good sign that the chiplet revolution is actually showing signs of coming to fruition.

Keith: 36:42

Yeah, definitely a good point, Chris. Actually, you triggered something. I want to take us down this rabbit hole just for a second. So, we're talking about advanced packaging and we're talking about thermal and power and all these chiplets being integrated and it's all to, of course, improve performance, reduce cost, but a lot of it is really we need to reduce the power, but it's really interesting to contrast that to the human brain, right? So the human brain uses about 10 watts of power. So it's orders of magnitude I mean logarithmic scales of magnitude difference and you get comparably you know, not equal performance. Obviously, the large language models are amazing, but the price you pay for it is immense in terms of power. Okay, so my point or my question is during your research, did you talk to any of the companies about biological computers, and is that something that you found throughout the research of the book? And if so, any thoughts or any comments you would have on that would be interesting.

Chris: 37:58

Well, I think in some ways, computing has been trying to learn from biology for a long time.

Chris: 38:05

If you think, for example, of neural networks, it's now been decades in which computer scientists have been trying to think about how do brains work and try to replicate them in certain ways, and so you could argue that the entire boom in AI today stems from a process of trying to learn from biology that was kicked off some time ago now. But I think you're right that there are certainly some interesting research ideas for how we can use biology to compute more effectively in the future, dna-based computing, for example. My sense is we're still a very, very long way away from any practical implementation, and the reason is right now the cost that it's so easy to innovate on silicon relative to any other material we know, so easy to innovate with ones and zeros relative to any other type of programming, that we know, such a developed labor force around this particular ecosystem and it's going to be really hard to move to something new. And so, even if I look at a sphere like quantum computing, which is, I think, substantially more developed than biological-based computing paradigms today, although still in its early stages, to me it's going to be a long time before we're really deploying quantum computing at any sort of meaningful scale, if indeed we ever do. And that's just because we've built up so much extraordinary expertise in essentially silicon-based digital computing. It'll be hard to replicate in any other sphere.

Keith: 39:32

Yeah, the economies of scale that you get from just silicon manufacturing dwarfs everything that I've seen. And again, just as a sort of an aside, back in the late 90s, early 2000s, when mobile phones were blowing up, there were different technologies that were used inside the mobile phone. There was silicon logic, but there were the RF radio transmitters and receivers. They benefited by using gallium arsenide materials, which had a much higher performance, but the problem was the cost. So, if you could continually improve the silicon, even though the performance wasn't as good as the gallium arsenide, eventually it would catch up, and that's exactly what happened. So, to your point, when you have an entire industry that's focused on, let's say, silicon as the main material, it's very difficult to get off of that, unless you run into some brick wall that just forces you to move.

Keith: 40:43

Okay, so a couple more. You obviously get interviewed a lot. Your book is outstanding and it's getting a lot of play in the semiconductor arena. So, what one question is it that you would like to be asked, or that you think is extremely important but don't get asked, or that is asked infrequently?

Chris: 41:05

Well, I think there's lots of discussion in the public sphere today about semiconductors, at least a lot more than there historically has been about supply chains, about materials, about the roles of semiconductor and AI. I think that the factor that most people underestimate on the question that I don't get asked nearly enough is about how do you actually build a workforce of people that is capable of pushing forward technological boundaries and applying semiconductor knowledge to lots of different spheres, and to me, that's really critical. If you go back historically, certainly we've had Moore's law giving us better and better computing and making our data centers, computers, better and better, but we've also had the application of these techniques in new spheres. So you mentioned mobile phones. Well, it wasn't guaranteed that we would have people who understood how you could take semiconductors and use them to make mobile phones possible. That was only because we had a large number of people engineers working on semiconductor issues, and some of them were exploring communications technologies, and enough of them started playing around in the sphere to develop these types of capabilities.

Chris: 42:14

And so, to me, part of the challenge we face is taking the technologies we have or are developing and learn how to deploy them in lots of different spheres of life, because my suspicion is that we've only actually begun to scratch the surface of the number of ways we could use semiconductors to do things more efficiently. We've figured out how to do it on computers and our phones and in cars increasingly. But if you look at the number of devices today that have semiconductors relative to the number of devices that 30 years ago had semiconductors, what you find is that we still are learning lots of new ways to use semiconductors and I suspect there's a lot more of that to cover the future. But it's going to depend not only on technology but also on the know-how, the expertise in the workforce to actually deploy this in many different segments of the economy.

Keith: 43:02

So it's interesting that you mentioned the know-how and the workforce. So, that actually generates another thought/comment. The large language models like ChatGPT, which are equivalent to at least a senior expert in any domain, material science, physics, economics, history, you name it. If you have an agent or a series of agents like that, that can assist you in your job, your career, your education, whatever it may be, it strikes me as maybe we're not in such a bad shape as we were in the past because we have now at our fingertips really experts in every major discipline that could be needed. So I'm just wondering or curious on your thoughts on that, do you see it similarly or do you disagree, that's not really going to be helpful. Just curious to hear what you think on that.

Chris: 44:07

No, I think that's right or at least I hope that's right that we're all going to have AI powered assistants and advisors helping us make decisions and helping us become more productive in what we do.

Chris: 44:21

I think we're still in the early stages of actually taking advantage of the capabilities that we see, for example, in Chat GPT. I think if you did a survey and asked what share of the American workforce, for example, was seeing meaningful productivity increases today from Chat GPT, I think it'd be a very, very small share. Instead of a couple pretty specific spheres maybe software programming you're actually seeing meaningful productivity increases, but I think for most people, you'd see less than 1% productivity change today. I think what we're going to need is the deployment of these tools and curated ways for different use cases. We're already beginning to see that from certain companies, but my guess is that this will take some time to find the right way to deploy these tools and make them actually practical to make workers more productive, which is certainly what we're going to need for to really capitalize on what AI can provide.

Keith: 45:16

Yeah, definitely, I think the privacy restrictions and the data privacy just concerns are also, let's say, thwarting the adoption rate. So, a lot of companies are actually pushing back against that until there are better guidelines and better laws in place. But it will be interesting as that becomes more prevalent in society. So, your book, Chip War, I think undoubtedly it's educated, it's obviously influencing the entire semiconductor industry. What do you hope will be the lasting impacts or the takeaway for your readers?

Chris: 45:58

Well, I wrote the book to educate the typical person who was sort of like myself before I started researching didn't really understand why semiconductors were so important or how just mind-bogglingly complex they are to actually produce. So my goal is that the rest of the world appreciates more fully what the industry produces and that we continue investing in the types of research and development that will make ongoing progress possible. Because to me, to go back where we started the conversation, there's really been no force more transformative than our extraordinary advance in computing power that chip have made possible. And so I think once you realize that fact, you also realize that we've got to do everything we can, collectively, as a society and individual companies, to keep that progress going, because so much depends on it.

Keith: 46:49

Yes, and with that, Chris, I want to say I believe your vision of the outcome of that book is definitely having that effect. I'd like to thank you personally. The book had a big influence on me and also to the rest of Advantest, so with that I'd like to thank you again for coming on Advantest Talks Semi.

Chris: 47:11

Thank you.