No Priors Ep. 13 | With Jensen Huang, Founder & CEO of NVIDIA

1. 0:00 – 1:26

   Introduction

   1. SGSarah Guo0:05

      Thank you for doing this with us, Jensen.

   2. JHJensen Huang0:07

      Delighted to do it, Sarah. Delighted.

   3. SGSarah Guo0:10

      Why don't we start at the beginning? Uh, you worked at LSI and AMD before starting a company. How did that happen?

   4. JHJensen Huang0:17

      They gave me a job. (laughs)

   5. EGElad Gil0:20

      (laughs)

   6. NANarrator0:20

      (laughs)

   7. JHJensen Huang0:20

      Let's see, uh, let... I was at Oregon State University, and, uh, uh, it was, um, a campus company day, and, uh, I interviewed at a lot of companies. And, and, um, uh, uh, two companies really, really, um, uh, connected with me. I, I love designing chips and designing computers, and, uh, at the time, in our lab, in the computer science lab, there was a, a poster of a 29,000, um, 32-bit, uh, CPU from AMD. And, uh, uh, it... you know, I always thought it'd be kinda cool to build that. Uh, on the other hand, uh, there was a- another company that was a startup at the time, uh, built by, uh, one of the legends of Silicon Valley, Wulf Corrigan. And, uh, they started a company to design chips using software, uh, to design chips not by hand, but by, by, uh, uh, using programmable, uh, logic. And, and you would describe it in language, and it would, it would synthesize it to chips. And of course,

2. 1:26 – 4:58

   The early days when Jensen Co-founded NVIDIA

   1. JHJensen Huang1:26

      uh, I chose to go to AMD. Uh, it turned out, it turned out that I went there to design microprocessors and, and, uh, my lab partner... not my lab, but my office mate, uh, ended up going to LSI. And, and she, uh, uh, insisted that I, I go there, uh, after I was there and after, after she went there. And, and, um, uh, the LSI team, uh, said, "Hey, we were recruiting this kid from, from Oregon State, and we really wanted him to come work at LSI Logic." And, and it turned out to have been her office mate. And so they all reached out to me, and, and I decided to go there because, uh, it was at the beginning of the EDA industry. It was at the beginning of, uh, designing, uh, chips using computers. And, uh, it was probably one of the... one of the best things that's ever, ever happened to me. And, um, uh, it was in the beginning of the, the, uh, ability for every company to build their own chips, and it's the reason why, uh, I met some really great computer architects. Uh, uh, Andy Bechtolsheim was the founder of Sun. I got to work with, uh, uh, a bunch of great architects at Silicon Graphics and, uh, Jon Rubinstein, who, who, um, uh, was at a company called Dana Computer, who became the vice president of engineering for Apple. And, and so I... uh, in, in a very early age, I got the opportunity to work... And then, of course, uh, the two founders of NVIDIA, uh, Chris Malkowski and, uh, Curtis Payne, myself. And, and so I got a chance to work with some really amazing computer architects, and I learned a lot about, uh, about building computers with chips. And so that's, you know, my early days.

   2. SGSarah Guo2:55

      And you were a star at LSI with your co-founders. At what point did you know, "I have to start a company"?

   3. JHJensen Huang3:02

      Uh, it wasn't my idea. It was theirs. Uh, Chris and Curtis wanted to, wanted to leave Sun. Uh, they had their own reasons. And, uh, I was doing really well at, at LSI Logic, and, and I enjoyed my job, and we had two kids, Laurie and Ian. And, um, I... and, you know, just like, just like you, they, they wouldn't stop hounding me.

   4. EGElad Gil3:21

      (laughs)

   5. JHJensen Huang3:23

      And, and they said, "Hey, you know, we, we wanna start this company, and we really need you to come along." And, and, uh, and I told them that I really needed to have a job, and, and, um... and so anyways, uh, they needed to figure out what to do. Uh, at the time, the, the S- the valley was, was, uh, um... the way of designing computers was rather s- rather split between, uh, general purpose computing, uh, versus using accelerators. And, um, about 99% of, of, uh, of, uh, the valley was, uh, believed in general purpose computing, and about 1% believed in acceleration. And, um, and, and for 25 years, 99% was right. So we s- we, uh, we decided to start a company on, on accelerated computing. Um, and, you know, at the time, the only thing you could really do with accelerated computing is, is, uh, find applications or find problems that were barely solvable or unsolvable by general purpose computing. So then that's k- kinda what we dedicated our company to do, to solve problems that normal computers can't. And if you, if you, uh, follow that mission to its limit, um, it led us to, uh, self-driving cars. It led us to robotics. It led us to climate science problems, uh, digital biology, and of course, you know, one of the most famous ones is artificial intelligence.

   6. SGSarah Guo4:43

      Uh, so you were working on this huge set of applications before the, you know, current ra- wave of artificial intelligence. What was the original, um, technical advantage of NVIDIA in, in artificial intelligence, and when did you begin to realize that,

3. 4:58 – 10:42

   Why NVIDIA started to expand its aperture to artificial intelligence use cases

   1. SGSarah Guo4:58

      um, this was gonna be a por- important use case for you guys?

   2. JHJensen Huang5:02

      Uh, so we, we had... we had, um, uh, expanded the, the flexibility of our, of our, um, of our accelerators to, to be more general purpose, and we invented a, a new computing model, uh, called CUDA. And, um... we're doing this podcast, like, at 4:00 or something like that in the afternoon. It was like, at the lowest point of energy. Isn't that right?

   3. EGElad Gil5:28

      Yeah.

   4. JHJensen Huang5:30

      So this, this is... we need some...

   5. EGElad Gil5:32

      That's why we need-

   6. JHJensen Huang5:32

      We need some nerds.

   7. EGElad Gil5:34

      ... nerds.

   8. SGSarah Guo5:34

      We need nerds.

   9. JHJensen Huang5:34

      That's right, we need some nerds, guys.

   10. NANarrator5:35

       (laughs)

   11. SGSarah Guo5:36

       Thank you, nerds.

   12. EGElad Gil5:38

       Now with gummy clusters as well, so it's very exciting new technology.

   13. JHJensen Huang5:41

       We need, we need, we need some energy.

   14. EGElad Gil5:42

       (laughs)

   15. JHJensen Huang5:42

       We need some accelerated... we need some accelerated computing in our life.

   16. SGSarah Guo5:45

       (laughs)

   17. NANarrator5:45

       (laughs)

   18. JHJensen Huang5:46

       And my, my voice is hoarse. (clears throat) So we wanted to... we wanted to make our, our, um, our graphics processors more and more general. And the reason for that in the beginning was because some of the effects that we had to do related to general purpose image processing, post-effects. You render an image, and you do post, post, um, image effects.... and other applications, of course, we wanted to bring the scene to life, and so we had to do physics processing. Um, in order to do physics, you have to do particle physics, fluid dynamics, so on and so forth. And so we, we expanded the aperture of our, of our accelerated computing platform to be more, and more, and more general purpose. The problem with general purposeness is that the more general purpose you are, the less, um, acceleration you get in any particular domain. And so you've got, you've got to find that, that, that line really, really carefully. And that's one of the gifts of our company, to find that line between, on the one hand, every single generation bringing enormous amounts of acceleration well beyond what CPU could do, um, to the application. And so, if you become too general purpose, you're like, just like a CPU. How can you accelerate a CPU with a CPU? And so you... So you have to find a way to f- walk that line. On the other hand, if you don't expand the aperture of the applications that you serve, the R&D dollars that you're able to generate wouldn't be enough to stay ahead of the CPU, which had the largest R&D budget of any chip on the planet. So if you think about this problem, it's actually really nearly impossible, because you have a small application, let's call it, you know, a billion-dollar market at the time, and, and out of that billion-dollar market you're investing 115- $150 million a year. Out of that $150 million a year, how do you keep up with a few hundred billion dollar industry? It's not even sensible. And so you have to find that niche very, very carefully where $150 million would accelerate this particular application abnormally and insanely. And then over time, you could expand your application space so that it's, goes from a billion dollars, to five billion dollars, to ten billion dollars, so on and so forth, without falling off that cliff. That is the fine line that we w- we walk, and... And so we kept expanding the, the general purposeness, and it led us to, uh, molecular dynamics simulation, which is what this image seems to look like. And, and, um, uh, seismic processing was another, uh, industry. And, uh, just slowly but surely, uh, we, we expanded our aperture. But one of the things that we did well was to make sure that, that irrespective of whether somebody used our platform for general purpose computing, accelerated computing, we always maintained the architecture compatibility. And the reason for that is because we wanted, uh, a platform that would attract developers. If every single NVIDIA chip in the world was incompatible, then how would a developer be able to pick one up, even if they learned that, that CUDA was going to be incredible for them? How would they pick that up and say, "I'm going to develop an application that's going to run on that"? Wh- h- which chip would they, would they have to go figure out? And, and nobody could figure that out. And so we said, "If we're gonna... If we believe in an architecture, and we want this to be a new computing platform, then let's make sure that every one of our chips, uh, perform exactly the same way, just like an x86, just like ARM, just like any computing platform." And so for the first five, ten years, you know, we had very few customers for CUDA, but we made every chip CUDA-compatible. And you can go back in history and looked at, looked at our gross margins. Um, it started out, it started out poor, and it got worse, you know? (laughs) So... Uh, because we were, we were in a really competitive industry and, and we were still trying to figure out how to do our job and build cost-effective things, so... So you know, i- i- it was already challenging as it is, and then we layered on top of this this architecture that was called CUDA that had no applications that nobody paid for.

   19. EGElad Gil9:43

       Yeah.

   20. JHJensen Huang9:44

       And so-

   21. EGElad Gil9:44

       It's kind of amazing, because now when I talk to people in the AI world in terms of one of the reasons that they really love using NVIDIA GPUs is because of CUDA.

   22. JHJensen Huang9:51

       Mm-hmm.

   23. EGElad Gil9:51

       And then because of the ability to scale interconnect.

   24. JHJensen Huang9:53

       Yeah.

   25. EGElad Gil9:53

       And so you can really, like, highly parallelize these things as well, which you can't necessarily do with other approaches or architectures that are in the market today.

   26. JHJensen Huang9:59

       Yeah, and s- so this, this computing platform, Elad, it, it's, uh, i- i- it's strange in a sense that it does, it performs these miraculous things, um, and, and, uh, we carried it to, to the world on the backs of GeForce, which is a gaming card.

   27. EGElad Gil10:12

       Mm-hmm. Mm-hmm.

   28. JHJensen Huang10:13

       Uh, the, the first GPU that Geoff Hinton got for his lab-

   29. EGElad Gil10:17

       Mm-hmm.

   30. JHJensen Huang10:17

       ... you know, uh, E- Elad would tell you that, that, uh, Geoff came in and said, "Here's a couple of, uh, uh, GPUs."
4. 10:42 – 13:52

   The moment in 2012 Jensen realized AI was going to be huge

   1. EGElad Gil10:42

      and sort of mining, and then in the context of AI. And it seemed like those were the two markets where a bunch of people were organically just adopting you. Was it... Were you marketing to those communities? Was it just people started realizing that-

   2. JHJensen Huang10:53

      That's the beauty-

   3. EGElad Gil10:53

      ... they needed linear algebra and... (laughs)

   4. JHJensen Huang10:54

      That's the beauty of a computing platform, right?

   5. EGElad Gil10:56

      Yeah.

   6. JHJensen Huang10:56

      Uh, in the beginning, you have to target the applications, and in the beginning we did. Uh, one of the first applications was NAMD.

   7. EGElad Gil11:02

      Mm-hmm.

   8. JHJensen Huang11:03

      Uh, seismic processing. Um, uh, it was, uh, both of them, both of them are, are, are... Uh, one of them is kind of particle physics, the other one is, uh, image processing, if you will, so... And so inverse physics, if you will.

   9. EGElad Gil11:15

      Mm-hmm.

   10. JHJensen Huang11:15

       And, and so, uh, one particular domain, uh, you know, we just, we just went out to hire, uh, to, to, um, uh, research, we went to scientific computing, um, centers and, and we said, "What kind of pro- problems are just beyond your reach?"

   11. EGElad Gil11:28

       Mm-hmm.

   12. JHJensen Huang11:29

       And, uh, um, the list of applications are, include-

   13. EGElad Gil11:32

       Yeah.

   14. JHJensen Huang11:32

       ... quantum chemistry and quantum physics and, you know, so on and so forth.

   15. EGElad Gil11:35

       What was the moment when you said, "Wow, this AI thing is really important for us?"

   16. JHJensen Huang11:39

       Uh, it happened, it happened, uh, uh, around 2012, I guess, and it was because, uh, simultaneously, Andrew Ng, uh, reached out to Bill Dally, our, our chief scientist, um, to, uh, work on a way to get the neural network model that they were working on-

   17. EGElad Gil11:56

       Mm-hmm, mm-hmm.

   18. JHJensen Huang11:57

       ... uh, onto GPU so that they could, instead of using, uh, thousands of, uh, CPU servers, they could use a few GPUs, uh, to, uh, to do training. So, that was one.... simultaneously, almost— uh, l- simultaneously, uh, Geoff Hinton reached out to us, and we started-

   19. SGSarah Guo12:13

       Hm.

   20. JHJensen Huang12:13

       ... hearing about that, and, uh, same thing was happening with Yann LeCun in his lab, and-

   21. SGSarah Guo12:17

       Mm-hmm.

   22. JHJensen Huang12:17

       ... and so simultaneously in several different labs, we're starting to feel that there's this, this, uh, this neural network-

   23. SGSarah Guo12:23

       Mm-hmm.

   24. JHJensen Huang12:24

       ... uh, emergence that, that is ... and, and that attracted our attention.

   25. EGElad Gil12:29

       Yeah, I guess 2012 was also the year when AlexNet came out.

   26. JHJensen Huang12:31

       Yeah. Right.

   27. EGElad Gil12:31

       So it felt like that was a year of transition for deep learning in general, in terms of really ... that was the moment in time, at least, that I remember thinking, "Wow, this, this really exciting wave of AI coming."

   28. JHJensen Huang12:40

       Yeah.

   29. EGElad Gil12:40

       And then I feel like for 10 years nothing really happened for startups, but a lot of incumbents started adopting this technology at scale. Yeah.

   30. JHJensen Huang12:45

       Yeah, we started feeling it. We started hearing about it before that, and then ImageNet kind of-
5. 13:52 – 17:48

   How we’re in a broader platform shift in computer science

   1. JHJensen Huang13:52

      I mean, just, just if it was just computer vision, we could've used it for all kinds of int- interesting applications, like self-driving cars and robotics, and, and we did. Uh, but w- we observed that this might be a new way of writing software altogether and asking ourselves, "What's the implication to chip design, system design, interconnect the algorithm, the system software?" Uh, to, to really reason about not just, "Why is this exciting, why was it so effective," which was impl- i- it was ... j- that alone was plenty, miraculous.

   2. SGSarah Guo14:26

      Mm-hmm.

   3. JHJensen Huang14:26

      That ImageNet, uh, without, without, uh, a specifically any human engineered algorithm would, would reach the level of effectiveness, you know, compared to 30 years of computer vision al- algorithms, uh, overnight, you know, and it wasn't, it wasn't by a small amount. And so the first question, of course, is, why is it so effective, and was this going to be scalable? And if it was gonna be scalable, what's the implication to the rest of computer science? What, what problems can't this universal function approximator, if you will, that can solve problems of dimensionality extraordinarily high, and yet you could, um, uh, learn the function using enough data, which, at the time, we were starting to believe we can get plenty of, and to, to, uh, systematically train this model into existence? Because, you know, it's, it's, uh, you train 'em one layer at a time.

   4. EGElad Gil15:23

      Can you talk a little bit more about ... I've, I've heard you be very articulate in terms of how you view this as a broader platform shift, just even in terms of, like, how pages are served versus, you know, generated or other aspects of that. Could you talk a little bit more about, you know, what's really happening right now more broadly in computer science with this shift to AI?

   5. JHJensen Huang15:39

      Yeah, so you fast-forward now a decade. The first, the first five years was, was by reasoning the impact of computer science altogether. Um, th- at the same time, we're developing, uh, new models of all kinds, right? And so CNNs to ResNets to RNNs to LSTMs to, you know, all kinds of new models, and, and scaling 'em larger and larger, making great strides in perception models particularly. And of course, the, uh, the transformer was a big deal, BERT was a big deal, um, a- all, all of you know that story r- well, and, um-

   6. SGSarah Guo16:12

      Did you guys see, like a, like a step change in, in volume growth, right, uh, with transformers and, and BERT and such? Um, because it feels like having a architecture and me- like, an attention mechanism that allowed for scaling of these models really was also a kickstart in the industry.

   7. JHJensen Huang16:29

      Well, the, the ability for you to learn, um, patterns and relationships from spatial as well as sequential data, uh, must be an architecture that's very effective, right? And so, so I think on its, on its first principles, you, you can kinda think transformer's gonna make ... it's gonna be a big, big deal.

   8. SGSarah Guo16:49

      Mm-hmm.

   9. JHJensen Huang16:49

      Not only that, you could train it in parallel, and you can really scale this model up. And so that's very, very exciting. Um, I ... and so I, I think that, that when transformers came, first came out, uh, we realized that, uh, that there's a model now that overcame the limitations of RNNs and LSTMs, and, and, um, uh, we can now learn sequential data in, in a very large way. So, that was very exciting. Uh, BERT was very exciting. Um, uh, we trained some of the early language models ourselves, and th- and we saw very good results. Uh, but it, it wasn't until, it wasn't until, um, uh, uh, the combination of, uh, reinforcement learning/human feedback, uh, wasn't ... and, and of course some of the breakthrough work that was done with retrieval models, um, uh, dialogue managers that, that does the g- guardrailing. Um, s- it wasn't until some of, all of those kind of pieces started to come together that, of course, that we all enjoy ChatGPT. And, and, Eli, the, the point that you're trying to make

6. 17:48 – 18:09

   His vision for NVIDIA’s future lines of business

   1. JHJensen Huang17:48

      is, is, um, uh, the, the observation that computer programming has now, uh, been completely disrupted, that for the very first time in the history of computing, the language of programming a computer is human.

   2. EGElad Gil18:01

      Mm-hmm.

   3. JHJensen Huang18:02

      You know? Th- any human language. And, and, uh, it doesn't even have to be grammatically correct.

   4. SGSarah Guo18:07

      (laughs)

   5. JHJensen Huang18:08

      And, and it's, it's,

7. 18:09 – 25:41

   How NVIDIA has two motions: Shipping reliable chips and solving new use cases

   1. JHJensen Huang18:09

      uh, fairly in- incredible that, that, um, anyone can program a computer now.

   2. EGElad Gil18:14

      Mm-hmm.

   3. JHJensen Huang18:14

      And so that's a, that's a, that's a big deal. Uh, the fact that you program it differently, it writes different applications.... um, what is the reach of this new computing, uh, computing model, uh, apparently quite, quite large, and it's, it's the reason why ChatGPT is the fastest-growing application in history.

   4. EGElad Gil18:30

      Mm-hmm.

   5. SGSarah Guo18:31

      We had, um, Alex Graveley, who is the chief architect for Copilot-

   6. JHJensen Huang18:35

      Uh-huh.

   7. SGSarah Guo18:35

      ... on the show as well, and, uh, his, his favorite... Like, obviously it's, you know, very powerful to have sort of, um, uh, sequential code prediction, but his favorite use cases of Copilot have been, like, people telling him that they don't code, but now they do.

   8. JHJensen Huang18:48

      I... Yeah, right.

   9. SGSarah Guo18:48

      Which is, which I think is very democratizing, as you said.

   10. JHJensen Huang18:51

       It's quite amazing that you could-

   11. SGSarah Guo18:52

       Yeah.

   12. JHJensen Huang18:52

       ... give, uh, um, ChatGPT a, a problem to solve, and it reasons through it step by step, um, but yet it, it arrives at the wrong answer on the one hand. On the other hand, you could tell it to write a program to solve the same problem, and it writes a program that solves the problem perfectly. And so, the, the fact that, that there's an application that, on the one hand, reasons and tries to solve a problem and does a fairly good job at it, it's almost there. Um, uh, on the other hand, it can write a program altogether to solve that same problem. You know, it's, it's, uh, you gotta really w- wrap your head around the implication of this, um, uh, yeah.

   13. EGElad Gil19:27

       So do you view it as, like, the future world is some form of machine sentience?

   14. JHJensen Huang19:31

       Well, uh, first of all, I don't even know what that word means to... Uh, uh, in a-

   15. EGElad Gil19:34

       Yeah.

   16. JHJensen Huang19:34

       ... in a technical way.

   17. EGElad Gil19:36

       Yeah.

   18. SGSarah Guo19:36

       Yeah.

   19. JHJensen Huang19:36

       Um, I'm, uh, you know, I'm fairly sure that I'm sentient, less so today.

   20. EGElad Gil19:42

       (laughs)

   21. JHJensen Huang19:43

       I'm, I'm kind of under the weather, but-

   22. EGElad Gil19:44

       That's what we have nerves for.

   23. JHJensen Huang19:44

       Yeah. Yeah, that's why I need-

   24. SGSarah Guo19:45

       Keep eating. (laughs)

   25. JHJensen Huang19:45

       That's why I need nerds to, to, to, to crank me up here.

   26. EGElad Gil19:48

       (laughs)

   27. SGSarah Guo19:48

       I'll try to.

   28. JHJensen Huang19:49

       Yeah, I know. I know, today was a tough day. No. Um, but, but, uh, I, I, I don't, I don't know. But, uh, do I, do I believe that, uh, we now have a, uh, a software that can reason through a problem, uh, for many, many types of problems, reason through a problem and solve... and provide a solution, uh, or a program to systematically provide a solution on an ongoing basis? The answer is yeah.

   29. EGElad Gil20:15

       Mm-hmm. Yeah. And then as you look forward to that world, how do you think about, um, where you wanna take NVIDIA's lines of business, but also, you mentioned in the past that NVIDIA has done things like train models, and you've done some really interesting things there. Is that gonna be an increasing part of what you do in the future, or are you mainly focused on the chip side? Or how do you think about that mix of helping to push forward some research as well as, you know, being the underlying platform for the industry?

   30. JHJensen Huang20:38

       Well, we're a computing platform company, and we have to go u- go up the stack as far as we need to so that developers can use it.
8. 25:41 – 31:39

   Why no one should assume they’re right for the job of CEO and why not every company needs to be architected as the US military

   1. JHJensen Huang25:41

      Investing in the future and being, um, sustainable now are, are not in conflict with each other. And, and so the, the, the challenge for, for all startup CEOs and for all CEOs is to find a way to, to be able to do what you believe in, the, the fundamental core belief of the institution, and to be able to afford doing it. That is the purpose of the company.

   2. SGSarah Guo26:08

      Mm-hmm.

   3. JHJensen Huang26:09

      And, and so you... The... It's part, it's part, um, conviction, it's part skill. You know, make me... Making money is not a, a matter of conviction. Making money is a matter of skill, and it's a learnable skill. And it took me a long time to learn it, you know, I'll, I'll admit that. Um, and, you know, I've been, I've been at this for 30 years, and, and, um, uh, for the first... For... Well, apparently, for the first 20 years, since you went back 10 years, uh, for the first 20 years, I was still trying to figure it out. And so... But it's a skill. Learning, learning how to make money, learning how to, how to, um, uh, uh, run a company, uh, efficiently, those are all skills, and the company has to develop the skills. And, and so, so I, I think the, the, um, the, the way that we, we ultimately do it, uh, is y-... We ask ourselves, "Do we... Do we really believe it or not?" And if we really d-... Believe in doing something, then it, it is the purpose of the enterprise, it's the singular purpose of the institution to go pursue its beliefs. And the rest of it is up to, um, all of the cleverness of the company and, you know, try to do our jobs well and build things that people want to buy and, um, try to make it as cost-effective as possible, make the company as efficient. Those are all skills. You know, the hard part, the hard part, uh, as, as it turns out, is not the skill part. Uh, you know, it took me a long time, but, but, um, uh, you know, a lot of companies know how to make money, obviously. And so, so the fact that there are more than one company that makes money suggests it's not that hard. You know, if somebody else can do it, Harcard can be.

   4. SGSarah Guo27:45

      (laughs)

   5. JHJensen Huang27:46

      (laughs) And, and so the... But to, to be singularly advancing, um, a new computing model we call accelerated computing, and we believe that someday, that on the one hand, accelerated computing can help us solve problems and tackle problems that normal computers can't, and it exposed us to all of these amazing applications, like digital biology, that I'm excited about today, like climate change, that we're excited about, like robotics and self-driving cars. If not for the fact that we're pursuing applications that were impossible with normal computers, why would we have discovered all of those things? You know, why would we have discovered artificial intelligence? Why, why would we be, um, the workhorse of large language models? Because large language models are barely possible. And if you are doing something that's barely possible, you call us, you know. We're the... You know, we're, we're the horse you call- (laughs)

   6. SGSarah Guo28:41

      (laughs)

   7. JHJensen Huang28:41

      ... to, to, to, uh, solve those problems. And, and so I, I love that aspect. I love the fact that we get to discover those future. Um, on the other hand, uh, w- we deeply believe that, that someday everything will be accelerated. And the reason for that is, is, um, very clearly that, that the CPU will run its course, and there's a limit to how far you could scale general purpose computing. And, and, um, uh, you'll always need it. You'll always need CPUs. But, uh, the, the type of applications that you... that we're, we're all gonna run, um, acceleration is really the best way forward. And so, so I, I... We just... A- at our core, we believe that from day one. 30 years ago, that's the reason why we started the company. And so it's the true conviction.

   8. SGSarah Guo29:25

      You have been, uh, enormously vindicated on this, you know, 30-year belief.

   9. JHJensen Huang29:30

      Mm-hmm.

   10. SGSarah Guo29:31

       Um, uh, you must have felt like, you know, that conviction challenged at some point in, you know, 30 years of running the company and learning the skills to run the company.

   11. JHJensen Huang29:39

       Mm-hmm.

   12. SGSarah Guo29:40

       What was the... I, I guess, what was the nearest death experience or the most concerned where you're like, "Maybe I'm not right," or has that ever happened?

   13. JHJensen Huang29:48

       That I'm not right for the job?

   14. SGSarah Guo29:50

       No, you're not right about accelerated computing and how important it will be.

   15. JHJensen Huang29:53

       Th- the second one's yes. (laughs)

   16. SGSarah Guo29:59

       (laughs)

   17. JHJensen Huang29:59

       (laughs) Uh, I, first of all, I, I don't think anybody, anybody should assu- assume that they're, they're right for the job. And so, so, uh, y- you should be gut-checking on that almost every day. And so-

   18. SGSarah Guo30:09

       To be clear, that wasn't the question, but-

   19. JHJensen Huang30:11

       But I'll, I'm more than happy to answer that.

   20. SGSarah Guo30:12

       Very humble of you. Yeah.

   21. JHJensen Huang30:13

       (laughs) ... answer that question. Um, uh, d- did I ever believe that, that was... it was wrong? No.

   22. SGSarah Guo30:18

       Mm-hmm.

   23. JHJensen Huang30:18

       I, I believe that accelerated computing is the... uh, absolutely the only way to solve problems that are impossible by definition if it's not-

   24. SGSarah Guo30:26

       Mm-hmm. Okay. Axiomatically.

   25. JHJensen Huang30:27

       Right? Mm-hmm.

   26. SGSarah Guo30:27

       Yeah.

   27. JHJensen Huang30:28

       And on the other hand, if you can solve problems that are impossible today, and someday you need that promp-... You need that application to be broad-based-... would accelerated computing be the best approach? The answer is yes. Yeah.

   28. EGElad Gil30:42

       When do you think the CPU hits its limits? You mentioned that, you know, eventually you think everything will move over, or at least, you know, big chunks of the future will move over. Is that five years away? Ten years away?

   29. JHJensen Huang30:52

       For certain applications, it happened-

   30. EGElad Gil30:54

       It's already there.
9. 31:39 – 32:57

   What’s next for NVIDIA’s Hopper

   1. JHJensen Huang31:39

      be generic. You know, every company in the world should not be built like the US military.

   2. EGElad Gil31:44

      Mm-hmm.

   3. JHJensen Huang31:45

      And in fact, if you look at every company's org chart in the world, they kinda look like the US military.

   4. EGElad Gil31:50

      Mm-hmm.

   5. JHJensen Huang31:51

      They, there, there's somebody on top and then it comes, you know, comes down and, um, uh, and yet, yet, uh, the number of direct reports of CEOs are very few and, and the direct reports of the people who are just learning how to manage-

   6. EGElad Gil32:05

      Mm-hmm.

   7. JHJensen Huang32:05

      ... first-level managers, are very large. It's exactly the opposite of how it should probably be architected.

   8. EGElad Gil32:12

      Mm-hmm.

   9. JHJensen Huang32:12

      You would think that the people that report to the CEO repor- requires no management at all and in fact, it's generally true. My direct reports are sophi- sophisticated, they're really talented, they're incredibly good at their job, they're excellent lead- leaders.

   10. EGElad Gil32:25

       Mm-hmm.

   11. JHJensen Huang32:26

       They have great business acumen, they have excellent vision. They're incredible.

   12. EGElad Gil32:30

       Mm-hmm.

   13. JHJensen Huang32:30

       Every single one of them. And, and, uh, why-

   14. SGSarah Guo32:34

       I guess that, that means, uh, you have more than like, you know, the sort of management book accepted, like, six or seven or whatever.

   15. JHJensen Huang32:40

       Yeah, I have 40-somewhat direct reports. And no one-on-ones, um, uh, no career coaching.

   16. EGElad Gil32:49

       Mm-hmm.

   17. JHJensen Huang32:49

       You know?

   18. SGSarah Guo32:50

       (laughs)

   19. JHJensen Huang32:51

       (laughs)

   20. EGElad Gil32:53

       (laughs)

   21. JHJensen Huang32:55

       "So, what would you like to do with your life?"

10. 32:57 – 35:08

    Durability of Transformers

    1. JHJensen Huang32:57

    2. EGElad Gil32:57

       Yeah.

    3. JHJensen Huang32:58

       Tho- those are conversations you have with, with new college grads and, and early career, and we love those conversations, of course, in, in helping them shape their career and mentor them, give them access to, to, uh, new experiences. Um, but, eh, but at the executive staff level, um, we're organized so that we can, we can pursue a whole lot of different things at the same time. However, one of the most important things about, about a software company is you have to understand, um, computer architecture and one of the most important thing about computer architecture is you can only afford one.

    4. EGElad Gil33:28

       Mm-hmm.

    5. JHJensen Huang33:30

       Just as, just as, um, uh, some of the largest companies in the world only have two operating systems, you know, the single largest company on the planet only has two. How is it possible that, that, that, um, uh, you know, so many companies have so many different computer architectures and they have, uh, seven or eight or nine instruction sets that they're keeping around? We have one instruction set. We have one computer architecture and we're super disciplined about that. And so where we need to be focused, we are. Where we, uh, allow for, for, um, uh, innovation and discovery, uh, at the senior level, uh, we allow that. And so, so I think the company's tapered, uh, and organized in a way that is consistent with the nature of our work, you know? So, that's the most important thing and, and that's probably the, the takeaway for, for what I've learned, um, uh, building our company is, there is no one generic architecture for every company. It should, it should fit, uh, the function of the company, its purpose, and then, of course, the leadership style of the, of the leaders.

    6. EGElad Gil34:34

       Mm-hmm. Yeah, I think that's a really important, um, note that most people don't really realize is that a company should almost be a bespoke structure supporting the CEO and their staff, and the way the company... What the company's delivering to customers versus it's always the same thing, and I think-

    7. JHJensen Huang34:47

       Exactly.

    8. EGElad Gil34:48

       ... that gets lost a lot.

    9. JHJensen Huang34:49

       Exactly.

    10. EGElad Gil34:49

        Yeah.

    11. JHJensen Huang34:50

        Yeah. There's some particular, um, chief that, that you need and a chief that, that you need-

    12. EGElad Gil34:55

        Mm-hmm.

    13. JHJensen Huang34:55

        ... and a chief this, that, you know, there's some chiefs that you do need.

    14. EGElad Gil34:57

        Yeah.

    15. JHJensen Huang34:58

        And, um, but aside from that, uh, you should start from first principles and architect something that, that makes sense for the, for the, the leader and, as well as the, the, uh,

11. 35:08 – 52:11

    What Jensen is excited about in the future of AI & his advice for founders

    1. JHJensen Huang35:08

       the function.

    2. EGElad Gil35:08

       Yeah.

    3. JHJensen Huang35:09

       Yeah.

    4. EGElad Gil35:09

       Yeah. When I, when I was at Google, they had the famous 80/20/10.

    5. JHJensen Huang35:12

       Mm-hmm.

    6. EGElad Gil35:12

       Where it was like 80% is core, 20% is, like, core adjacent/new stuff, and then 10% was hyper-experimental.

    7. JHJensen Huang35:18

       Mm-hmm.

    8. EGElad Gil35:18

       Do you have any frameworks or ways to think about that stuff or it's just kind of like let's see what organically is used in terms of this generic platform that we've built with CUDA and other things that are, you know, built in to help support a lot of use cases, and as they emerge, we, we say, "Okay, let's go support that new thing?"

    9. JHJensen Huang35:33

       Our str- yeah, I don't, I don't, h- I, I don't have any wise framework like that. Um, uh, there, there, there are a couple of things that, that our company has shaped, um, and structured to do. There's one part, uh, a very large part of our company is designed to, uh, build very, very complicated computers perfectly.

    10. EGElad Gil35:51

        Mm-hmm.

    11. JHJensen Huang35:52

        And so that's, that is, um, uh, one of its missions. Okay? And, and, uh, that kind of architecture, that kinda organization, uh, i- is a, is a, um, uh, invention and refinement organization. And then we have, we have, um, uh, a, a whole bunch of, of, um, uh, skunk works, if you will. And the reason for that is because we're trying to invent things 10 years out that we're not exactly sure whether it's gonna work or not and, and there's a lot of adaptation, a lot of pivoting, and, um, and so, so y- you know, our company actually has, has two different ways of working. One of them is rather organic, shape-shifting all the time. If a particular investment's not working out, we give up on it, move the resources somewhere else-And so that's the agile part of the company, and then there's a part of the company that's not rigid, but it's really refined.

    12. EGElad Gil36:45

        Mm-hmm.

    13. JHJensen Huang36:46

        And so these two, these two systems have to work side-by-side.

    14. SGSarah Guo36:50

        Can you talk a little bit about the H100?

    15. JHJensen Huang36:53

        Mm.

    16. SGSarah Guo36:53

        Next workhorse? And, like, what, um, what the most important innovations are and, like, what the, like, design and ship process for that looks like?

    17. JHJensen Huang37:00

        I would say the, the big breakthrough for Hopper is, uh, recognizing that quantization, the numerical quantization, the numerical formats, um, has, has a fair amount of, uh, innovation and, um, uh, uh, ability to, to reduce, uh, because it's, it's statistical in the first place. And now the question is, what kind of models could be created and trained and, and, um, uh, we believe that 8-bit floating point, uh, i- i- rather than, if you look at scientific computing today, 64-bit floating point. And so j- just by, uh, breaking up 64 into eight, you could increase the performance of an AI supercomputer just by a factor of eight, by not doing 64-bit. And so, so that's, that's almost a factor of, uh, if you will, a factor of 10 almost, in, in just a couple of generations, just by recognizing that 64-bit floating point wasn't necessary.

    18. SGSarah Guo37:59

        Mm-hmm.

    19. JHJensen Huang37:59

        And so, the, one of the big things is that. The second thing is, is Transformer. The Transformer engine is so universal and so useful that it's possible for us to design a, um, pipeline that is shaped for, uh, learning and, and, uh, inferencing Transformers. And so those are probably the two biggest things. Otherwise, it's the largest chip the world's ever made, it's, you know, the fastest chip the world's ever made, and, you know, super energy efficient and uses the fast memories that the world's ever made, and then we connect a whole bunch of these things together so that it's, it's fast and energy efficient. But those are all, you know, kind of brute force-y things, but the archi- the big architecture idea is, uh, FP8 and Floe- and Transformer engine.

    20. SGSarah Guo38:39

        And when you think about then, so that's the, you know, big project refinement part of the company.

    21. JHJensen Huang38:44

        Mm-hmm.

    22. SGSarah Guo38:44

        We think about the more agile piece, like, what's the impossible application you're working on today that's 10 years out you think is likely to be important?

    23. JHJensen Huang38:52

        M- tu-

    24. SGSarah Guo38:52

        I'm sure there are a ton of them, but...

    25. JHJensen Huang38:53

        The- th- there's a whole bunch we're working on that, that don't work at the moment-

    26. SGSarah Guo38:57

        (laughs) .

    27. JHJensen Huang38:57

        ... um, but, but I've got a lot of confidence it will work, okay? So for example, uh, uh, you know, autonomous driving is still, uh, making progress, uh, but I have every confidence that it will work. Um, uh, I have every confidence that a robotic foundation model will be discovered, and that, that, um, uh, through, uh, through, uh, expressing yourself, uh, using human language, um, uh, you could, you could, uh, cause a, um, uh, mechatronic system of almost different types of limbs and, you know, agility to be able to, uh, figure out how to bend itself, articulate itself, uh, to do a particular task. And, and, um-

    28. SGSarah Guo39:45

        What do you think the blockers are to that today?

    29. JHJensen Huang39:48

        Uh, I have no idea. Uh, I have no idea. Um, but I, I can't tell you. I'm just (laughs) -

    30. EGElad Gil39:54

        (laughs)

Episode duration: 52:11