Good evening everyone. I know that there was a line to get in and therefore we started a few minutes late here, but then our speaker promised that she's going to stay also a few minutes past six o'clock here. I want to welcome all of you to this edition of the presidential lecture series at Purdue University. And this will also be recorded because I know that some of the students couldn't get a ticket. It's free, but you have to get a ticket first. I heard it's very tough to get a ticket these days, it's like a lottery to Taylor Swift and so on. So to those who couldn't get a ticket, well, you know, don't worry, we're going to put it on video. In fact, we're going to also create a podcast on the audio side with a theme of technology-driven economic development here in the United States. And there's no better in my mind leader to help us inaugurate that podcast series and to get going with this presidential lecture than Dr. Lisa Su. Now Dr. Su is truly an immensely impactful, influential and visionary leader in the technology industry over the past many years. In particular after a sustained long leadership stint at IBM about 12 years ago, she became the CEO of AMD and if you would track AMD back 12 years ago, they were not doing very well. And Lisa turned it around multiple times with the evolution of cloud and mobility and AI these days. And I think you fact check me please, that 12 years ago it's about two plus billion dollar company AMD market cap and today's is 200 plus billion dollar depending on the day company. And also as many of you know Dr. Su had all her degrees undergrad master PhD from MIT and she is a real engineer as well and she is a member of the National Academy of Engineering, a member of the American Academy of Arts and Sciences, recipients of multiple awards including the SIA Robert Noyce Award and the IEEE Noyce Medal. So truly a unique and outstanding leader who spent some time with students earlier today at Birck and it is my distinct pleasure to welcome Dr. Lisa Su to Purdue.
Thank you. Thank you so much.
All right. Good evening everyone. Well, I just mentioned that it's hard to get a ticket and we're gonna have this on video and on audio as well. So that you don't have to hold your phone throughout the entire lecture. It's your freedom to hold your phone if you would like to. But you are quite a celebrity here at the land of the nerds. And we're very proud to have excellence at scale in what Purdue does. Now in particular we are going to talk a lot about in this actually whole series about the technology-driven economy and the shape through onshoring and AI to our future here but why don't we start perhaps with a bit of your personal decisions back when you had the choice upon getting your doctoral degree from MIT to go either academia or to industry. And you chose the industry and that's academia's grave loss. How did you arrive at that conclusion?
Well, first of all, President Chang, thank you for inviting me here. It's really a pleasure and honor to be here at Purdue. I heard there were extremely smart people here tonight, so we'll try to make sure that it's entertaining enough for you guys. Look, you know, I think my story is very much a story that is probably similar to many of you, right? I was born in Taiwan. I grew up in New York when my, you know, father immigrated to the United States for graduate school and he brought the entire family along. And yeah, I was a nerd. I think that's a pretty fair statement. I did quite a bit of research early in my career. And when I first was leaving school, actually I remember when I got my master's degree, I was so so in a hurry to actually, you know, go to work. And my parents, like many Asian parents, said, 'No, no, no, no, no. You have to get a PhD.' And I said, 'Really? Do I have to?' I wasn't convinced, but I did because I was a good girl and I did what my parents recommended. But this idea of what to do when you grow up, I mean, it's very hard to figure out what you want to do when you grow up. But what I actually thought is for those of you who are tenured professors, I thought that was just way too hard because if you want to be a tenured professor, you have to keep coming up with, you know, extremely new research, you know, all throughout your career. And if you go into industry, you can solve practical problems and build things. And I really enjoyed working as part of a team. And so I think my most favorite thing that I do, you know when I say you know what do I do? I still say I'm an engineer and what I like to do is build products and the fortune of being able to build products from very small teams to now products with a pretty a good-sized team is what I enjoy the most.
Well well an engineer at heart indeed. Now one of our colleagues Mark Lundstrom actually communicated with your former PhD advisor at MIT and we were told that you are being too humble. You absolutely had job offers to top research universities. I'm not going to name names you know which university's grave loss it was specifically but you you had your tenure track faculty offers clearly from outstanding institutions but you chose to go to industry. Well that's good news for industry because after IBM then to AMD now did I cite the numbers roughly correctly? I don't track stock market every day but
Close enough. Close enough.
So now from two plus billion to 200 plus billion if I can do math after five let's say 100x in 12 years that's not bad.
It's okay. The shareholders probably are feeling pretty good about 100x and it's not a startup right and you know going from valuation of 50k to 100 times bigger. So for a long-standing public traded company to go 100x in just a little over a decade and it's been a very busy decade with a lot of the technological transformations and what was your list of key moments perhaps in transforming AMD you had to compete against many other great companies right from Intel to Nvidia and Broadcom and you had to go through many rounds of technological transformation from cloud AI over the past 12 years and perhaps can share with the audience a few of those key transformational moments for AMD.
Okay, first I have to ask a question. So how many of you have used an AMD product in some way, shape or form? Either in your PCs, your gaming, your... All right, I like this crowd. I like this crowd. This is a good crowd. You can all stay now otherwise we'll have to swap to a different audience. Look you know a little bit of the history of AMD and actually the history of technology companies in general. You know AMD is 55 56 years old. We've had a long history. I think before I joined AMD I think we had a history of always being a challenger sort of the underdog. Really trying to ensure that there was you know very good compute out there. And you know when I joined AMD I thought it was like the best job in the world for me. It was like my dream job. And the reason is because I felt that we had an opportunity with an incredible set of assets you know team intellectual property you know kind of space where technology was becoming so important. But we had to make some different decisions and you know the real thing about any technology company is you must be focused on what you can be best at and for us our choice and you know Mark Papermaster my CTO Mark and I had a long history together we joined AMD about the same time and you know we had a passion for high performance computing and building you know sort of the bleeding edge compute out there and frankly at the time it wasn't the sexiest thing to do. I mean, it was all about mobile and everything was kind of going to smaller form factors and we fundamentally believed that the world would need more high performance computing going forward and we wanted to be best at that. And so, the beauty of sort of a tech company is you can start with an idea of what you want to do and you chart a road map and a path to get there. And we said it's going to take a while. We said hey it's going to take at least five years for you know for those of you who have experienced our products. I think Mark was very much the architect of our entire Zen product family. So the new PC processors, desktops, mobile servers and we really believe that this was become important and you know who would have guessed that today it's all about high performance computing. I think we had an idea that that would be the case but it's nice that that has come to fruition and you know we do spend sort of every day thinking about how do we push the bleeding edge of technology. You know being an engineer at training I remember and Mark Lundstrom and we were sort of in a similar field when we were working on device physics as a young researcher. I remember thinking the things that we are doing today like in terms of two and a half D integration, 3D integration, I thought it was virtually impossible. Like there is no way that this stuff could ever work in manufacturing at high volume and at scale. And somehow engineers are amazing when they're given a problem that they're trying to solve and there's a business reason to do it. It's not about how do you spend the most money. It's about how do you really figure out what are the most important problems that need to be fundamentally solved and somehow we find a way to do it. So that's the amazing thing about the world we live in and for all of you who are kind of starting out in your careers is like the things that you think are impossible are truly very very possible with kind of the right ambition and thought process.
Well, you mentioned that often AMD was viewed as the underdog and the market disruptor, the innovator. And what would you say to the was the talent business, the talents you would like to recruit, especially back 12 years ago? Yeah. And when you describe to say, hey, you have other job offers in the market as well. We like you to come to AMD. Well, yes, we are the underdog or proud of being the underdog when the focus on we're really good at focus on HPC and so on. What would you say to those talents you would like to recruit back then?
Well, you know, I would say a lot of people ask me why did I join AMD? And so when I'm asked that question, I say look, when you join a company, you want to join a place where you can make a really big difference. And you know, at the end of the day, you know, you train for this, you go to school for this, but you can work in many many different places. But the way we recruit people, the reason I came to AMD, the reason, you know, many of our team are here is because we wanted to operate at the bleeding edge of technology, but we didn't want to be like a cog in a big big, you know, in a big big pond. We wanted to truly, you know, shape where technology is going. And I think that's the freedom that you want to have. That's what I like to tell, you know, my team. You know the job of myself and our executive team is to ensure that our engineers have a chance to do the best engineering they can which is you know freedom to kind of dream about what you can do and the resources and the time to get it done.
Well, now AMD is doing so well and I know today you're visiting Eli Lilly and company and the Lilly CEO Dave Ricks is a Boilermaker and is actually now on the board of trustees here as well. Now you don't have to tell us anything you can tell us. But how's your conversation with Lilly?
Well first of all we thoroughly enjoyed the visit this morning to Lilly with Dave Ricks. You know, thanks for his invitation. And actually, the reason that I wanted to visit, frankly, and I brought a couple of my executive team is, you know, I'm a big believer in the way to really use technology at its best is when you bring together folks with different disciplines and different specialties and actually try to make 1 plus 1 is greater than 3. So the way I look at it is look our job is computing and you know we can talk about chips and software and systems all day long. We are not experts in life sciences. We are not experts in drug discovery, but we know that computing can be a huge sort of adder to that research environment and you know Lilly and you know Dave and his team have just tremendous tremendous capability in this area and so we're trying to make 1 plus 1 is greater than 3 and just in a few hour conversation we were able to identify areas that you know some of Dave's researchers would say hey these are areas whether you're talking about next generation manufacturing or you're talking about how do you use AI compute more effectively in the R&D process and we have a bunch of ideas of how that can possibly be done and that's what we do best. It's the idea of how do we make the compute foundation everybody needs more computing but the way you use it and how you really apply it to real world problems is what gets us excited.
Yes. Well, do you think that a generic whether it's training inference or a gigantic robotics ecosystem architecture suffices independent of the application domains or do you think that the application specificity should be an input to customization of chips?
I am a big believer in there is no one-size-fits-all in terms of chips and there's no one-size-fits-all in terms of models. You're going to have different types of chips for different applications. You're going to have different types of models. You have large models, medium models. You know, you have all of these foundational model companies that are competing today. But they're all slightly different in terms of what they're best at. And frankly, the winners, there's not going to be one winner. The winners are those who can, you know, take let's call it the generic foundational capability and quickly customize it and really tune it for the environment and the application and frankly I think that's where all of us are looking at the power of AI right? AI is a great foundational technology today but it is nowhere near as good as it can be if we apply it correctly.
Well, you know, I'm also doing commercial for my employer Purdue University here a little bit. That we have been working a lot on physical AI where the bytes of AI meet the atoms of what we grow, what we make, what we move. So agriculture, manufacturing, transportation industries and it is our belief and a bit of a bet to say that the domain specific models will be useful. And now you earlier announced last month a very big deal with Meta and some industry analysts are guessing that you're going to be competing very successfully against Broadcom and others and you mentioned multiple winners could exist in the customized chips industry. So maybe tell us a bit more about the arrangement between AMD and Meta.
Yeah, maybe if I give you a flavor of what are the key things that I think are important when you go through this current stage of AI. So again lots of excitement around ChatGPT and Gemini and all of these foundational model capabilities but we are still in such the early innings of how do you really take advantage of this capability and so one of our principles at AMD is to really build deep partnerships and as I the example I gave with Dave Ricks at Lilly, this very large deal that we announced with Meta is again how do we take our computing technology to help them unlock personal super intelligence. You know Mark Zuckerberg is very passionate about the idea that you can take AI into a much deeper personal experience and what do you need as the foundation of that? You need a lot of compute. These days when you think about these AI super factories you need to have relationships that go many many years in advance and we were talking about building gigawatt scale. I know Meta is building a plant not too far away from here here in Indiana or data center with the idea of installing the latest generation compute. So we're very excited to be partnered with some of the most important foundational model companies. Meta being one, OpenAI being another. You know, we're deeply partnered throughout the hyperscale ecosystem. But the whole idea is how can we make 1 plus 1 is greater than 3 by taking our expertise together with the software capability of giants like Meta and OpenAI and creating that next wave of AI that is much more efficient much more accessible and able to really solve the next generation of problems.
Well, you mentioned indeed Meta recently announced they are building a about 10 billion dollar hyperscaler...
A little bit expensive to build these days.
I heard we should talk a bit about their electricity usage as well. But that is indeed happening about half an hour away down the road from where we are right now. But if I could show a little map here, we could have that heart attack corridor. I hope you can all see that over there. Now that node who knows where that is on the well which node the node on the bottom there which city is that? Indianapolis and then we're right there West Lafayette Greater Lafayette and in between if you trust Google map that is exactly in between the middle point is the LEAP Innovation District in Lebanon Indiana and Lilly is building about 14 billion medicine foundry, three phases right there. And then the Meta that you just referred to location is going to be the neighbor in LEAP and Purdue now runs a heart corridor bus because we as of two years ago for the first time had a urban campus expansion of the main campus in Indy unlocking some unique opportunities for students and faculty. Now what we would like to visualize here is actually three industry sectors. There's the one health including medicine but as well as agriculture and animal vaccine. I see some of the great companies there listed. They're all located with their own factories or their own centers along the heart tech corridor. And then there's the whole AI tech stack from the applications to the models and to the hyperscaler to the chips and we'll talk a bit more about the electricity side as well and then there's the transportation rate at large including on the ground and in the air. So now there are many other industry sectors but our hope is to carefully curate many many tens of billions dollars of investment of private sector creating jobs workforce and innovation together along the 65 mile long America's heart tech corridor. So if you only want to write codes well it will be useful here but if you say I want to have technologies that I can touch the heart tech then this corridor will be an amazing place and of course it's Purdue everywhere along the heart tech corridor so it's golden black. Small advertisements at a small scale here so well we look forward to a lot of great growth here including as you pointed out the Lilly Meta side and now if we double click a bit on the AI tech stack. What do you think about the usage of electricity and what do you think about the roles that a small modular reactor in particular can play?
Yeah, there's no question when you look today at, you know, where what really drives AI, it is compute. Compute is the foundation for what unlocks, you know, I like to say compute equals intelligence. And so if you have the chance to install more compute to build more compute, you will but there are some physical limitations in doing that. Energy and power is one of the physical limitations. There's also things like chip supply and all those things. Anything that we can do about I would say seven or eight years ago it became absolutely clear that power would be one of the limiting factors in terms of long-term compute capability. So, it's something that, you know, we're very focused on as a chip company in terms of making our chips more efficient. We're always thinking about energy efficiency together with performance. And I think anything that we can do to generate power, you know, cleaner, faster, more cost effective will allow us to unlock the potential for more intelligence.
Indeed. Well, one of the companies, we didn't list all the companies with great footprints here along the corridor, but Caterpillar, you know, they've been very busy because of the gas turbines are used in powering data center, but we also a few months ago worked an arrangement with BWXT, one of the few American nuclear companies to explore the potential of a research micro reactor. Now, Purdue actually has the only fully digitized instrumentation and control digital small very very very small nuclear reactor for teaching and research purpose in the United States and is operating here on campus and you mentioned also this vision of what can be computed will be computed. So recently a mutual friend of ours Dario Gil talked about vision by liveness and you know I didn't ask Lisa directly I wasn't around but apparently he said something to the effect of everything that can be computed shall be computed one day. Okay all knowledge can be computed. AlphaFold for example as a Nobel prize winning success is a demonstration that all human knowledge can be computed and shall be computed. So I'm going to ask a contrarian view question then about this notion. Do you think that that is an efficient way to do it to compute in the current methodology and do you think that the monetization of the revenue generation or the business operation of the entire now intertwined AI ecosystem of all these major companies will be able to sustain on the business side this vision of all human knowledge shall be computed.
Yeah. Well, maybe I would say it a slightly different way because we don't necessarily have to have the extremes of should it be done this way or should it be done a completely different way. I think it's actually something different which is like we have really the luxury and opportunity to be at a time where we are so early in the development of a technology and like I say for all of you who are going to school today like I actually think you guys have a phenomenal opportunity to really be at the very very early stage of how technology will be applied. Now the way we're applying it today is not going to be the way that we do it five years from now or 10 years from now. I think AI is still so early in its innovation cycle. You know if you think about it we went from pre-training to inference to reinforcement learning to all of these various kind of new architectural techniques because we are learning right now. So to answer your question I don't think there's only one way of doing things. I don't think the answer is brute force like complete brute force because there are very very smart people out there. I mean this is one of the areas where I think it is quite amazing what startups have been able to do like we've been tracking many many startups in terms of just their innovation pace and their innovation cycle. And it used to be that these guys just didn't have the resources to get it done. But you actually don't need that many people and you don't necessarily need that much compute if you have a great idea. So my imagination is what we're going to find is that as we get more into the application stack like again Dave and I were talking about the healthcare industry which I am personally very passionate about because I view this as one of the areas that is just ripe for AI to disrupt. There's still a lot of fundamental science that needs to be learned like you can't ask the computer to go compute it because it doesn't know what the science is. So you actually still have to discover the science such that AI can be applied. But if we can do that in a way where you take the best experts from each domain and put them together in such a way that you can find the most efficient way to unlock intelligence then we all win. And so we should think about this as easily a 10 plus year cycle of learning. And you know each it's not even each year like every month you're surprised about how much the technology has moved. I mean I can tell you within the work that we're doing just internally with our partners like we are super excited about how much progress can be made in a short amount of time. But we're still so clear that we're nowhere near the pinnacle of what can be done. Like there's still so much evolution and learning in the technology which makes it a great place to be.
Indeed and you know AI was declared dead as a discipline decades ago one of my co-advisors Tom Cover actually told me and he wrote some of the fundamental papers back in the early 60s in neural network training fundamental limits that you know wow people don't think there's any hope in AI but then we saw that the tsunami of two things one is data and the other is compute capabilities empowered by chips such as AMD chips. I'm doing your commercial too. So thank you is mutual. So you know then say wow actually maybe there's a new world here. I remember once digressing a bit I asked Eric Schmidt this is almost 10 years ago now say hey you know the Google translator you know the translation of poems was not very poetic and he said well that's because not enough people care about poems to generate enough data for me to train you know give me enough data it will get much better now I remember another conversation and that's my predecessor Mitch Daniels he hosted a few years ago Gary Kasparov here in the same lecture series And one of questions was you know with Deep Blue you know it's very unfair in the sense that human brain's energy consumption is so much smaller what order of five order six orders of magnitude I don't know exactly compared to what Deep Blue was using so you give Deep Blue you know two apples a day to chew and see if they can beat you right now Kasparov's got a very classy answer he just said well doesn't matter so he's over it but I couldn't help but remember in the press release you and Meta had a few weeks ago, you actually counted by the gigawatts.
That's interesting because I thought, wait a moment, that's not the right unit. That's not a compute unit. That's a power unit.
To today, as I said, it is all about how much compute can you get in a fixed amount of power. Because when you look across the world, power is one of the largest limiters and it's how fast each country is building to satisfy that requirement and recognizing that it's a lot of power. We're talking about multiple gigawatts of power. So it's a lot of power.
Yeah indeed. Well you know one of our kids is doing AP calculus and she asked me well you know ChatGPT can answer my question about differential equation much better than you can. I said however do you know how much power that's using versus you know my brain you know I'm giving you the wrong answer but it was efficient. Well let me switch gear then a little bit to double click just on one layer and that is the specialty of your company in semiconductors. There's the design part and then there's the manufacturing and advanced packaging part. Well, you mentioned early on you and Mark Lundstrom others were looking at 3D or 2 and a half D integration heterogeneous integration and advanced packaging and if I remember correctly your PhD dissertation at MIT was SOI silicon on insulator MOSFET. So given this very special audience here, maybe you can tell us a bit more about your PhD dissertation.
I don't know if it's that relevant for this audience, but I'll say a few things. First of all, I had a chance to just spend a little bit of time at your Birck Nanotechnology Center and it is a phenomenal center. So congratulations really on everything that you have here. It is truly state-of-the-art. And it's really nice for me to see because you know it's funny like the world goes in waves right? So when I graduated long long long time ago I think hardware was still kind of okay and then we went through actually quite a period of time where software was a much more interesting discipline for people to be in and it's kind of nice to see there's a return to the hardware world but the example that I will give you is you know many people in semiconductors talk about Moore's law and you know Moore's law every 18 to 24 months you're supposed to double your productivity and capability and the entire world has been professing that Moore's law should end so my PhD thesis at that time which was I don't know this is like 1994
was how can we possibly extend Moore's law using silicon-on-insulator technology. But again, as I said earlier, many of these things look like they would be extremely hard to put into manufacturing. And what I have been super impressed by is just how much the technology has evolved. Like people used to say, I mean Mark you remember this right? People would say, hey semiconductors is a mature field, like no more innovation. You can't push lithography anymore. It's all about cost and nothing can be further from the truth. I mean the semiconductor industry has seen tremendous growth over the last four or five years and you really see not only do you see a lot of innovation, you actually see an acceleration of innovation. Like people are actually doing even more interesting things. There are different ways. There used to only be one way to improve performance, that was pure scaling. And now because scaling has slowed down there's so many other ways to do it, including 2.5D and 3D integration, including system level architecture, including obviously a lot of accelerators, you know whether GPU accelerators, ASIC accelerators. And I'm a believer in the world is going to be very heterogeneous. Like there is no one-size-fits-all. You're definitely going to optimize for the application. I think what this time really favors is those people who actually know what you're trying to accomplish in the application and also understand a little bit about the silicon capability because that is where you get to optimize across the stack. And I think it's probably as I look through my career the most valuable sort of asset that you get is you get to be an expert in things. I was an expert in device physics. I'm sure some of you are experts in a whole different set of fields. But the ability to kind of stitch together across things that you are not expert in but you can learn just enough so that you can connect the dots is actually the most fun part about it because it is about connecting the dots, right? We are all solving incredibly difficult problems that require multiple disciplines to come together to really make that next big leap forward.
Well, you mentioned that indeed Moore's law has been proclaimed dead many times and even 20 years ago. That's right. I remember when I was a grad student, oh don't go there because you know you have heard that Moore's law is dead. I say okay what's Moore's law again? Oh that very famous law but as one would say, you know, its death and demise has been grossly exaggerated and it's still alive and you just have to think maybe outside of the boxes. They said, Moon, take a look at this, you know, at this rate by the time you graduate they'll be splitting atoms, you know, that never have enough yield to do that. Well we're not splitting atoms to make it work but there are many other degrees of freedom. Now you mentioned also this hardware software, you know, there's a lot of engineering but also business architectural choices in how to divide and conquer and think about the interfaces, right? There's the hardware software, shall we integrate the operating system with the underlying chip or not? And there is also the different layering models in other engineering disciplines. There's a famous one in semiconductors and that is the bifurcation to say some companies will be fabless and do design and others are going to focus on making the chips. And for example SK Hynix is building this high bandwidth memory, the market leader in the world. Advanced packaging is part of manufacturing process here. And now what do you think about that bifurcation? Because some companies tried to do both and some companies made a choice, architectural decision years, decades ago to say I'm going to focus on this and let others focus on the other part.
Well, it is very much a business model choice and the business model choice has to do with capital intensity. It has to do with what you think you can be best at. Certainly we are a fabless company in the sense that we design semiconductors and we manufacture the vast majority of our chips at TSMC in Taiwan and also here in Arizona. Manufacturing is a discipline that you just need to be really, really expert in and it is not easy to be expert in so many different areas and so certainly we've made a choice. What is important is in the world that we're in today, resiliency in the supply chain is also very important. It's nice if you have great ideas but if you can't scale it and run it in volume production then it doesn't really help. And so a lot of what we think about is not just manufacturing the wafers but really these are very complex supply chains that now come into place. It's about wafers, it's about memory, it's about the system components that come together and it just says that these are all things that need to be considered. There's no one right model. Models are chosen for different choices. But I am a big believer in companies need to know what they're best at and you want to be super clear at when you wake up in the morning what are you trying to do. And for us it's about building high performance computing and AI at the bleeding edge.
What do you think about the architectural division between memory and logic?
It is very, very important to bring memory and logic closer together. The idea that a lot of system level performance is unlocked when you bring, you know, sort of how you operate on the data is very important. Again this is an area that's ripe for innovation because there's lots of different ways to put the systems together.
Now another commercial break here. I would like to just advertise that Purdue Semiconductor Industry Partnership research innovation as well as education including the semiconductor degrees program that does teach across all the dot connecting including design and manufacturing, advanced packaging, software, hardware, logic, memory. Now by the way whether you do it or not yourself it's very, very important to understand.
To know. Yeah, exactly. And we want our students to have hands-on experience doing that.
