On April 17, NVIDIA CEO Jensen Huang joined Cadence for a conversation with Cadence CEO Atnirudh Devgan to discuss the critical role of AI and accelerated computing in shaping industry megatrends and how NVIDIA and Cadence are partnering to drive transformational change in EDA, SDA, digital biology and AI.
Huang said AI is leading a quantum leap in computing power and is the epitome of computing transformation. Generative AI can increase computing power orders of magnitude higher than accelerated computing, and AI will revolutionize data centers, robotics/autonomous driving, and life sciences.
Regarding energy consumption, Huang believes that while AI consumes a lot of computing power, AI will ultimately save society in energy consumption by optimizing its design, and AI will be an important tool to combat climate change.
Regarding business management and R&D, Huang said AI tools will free engineers from heavy lifting and focus on innovative design and productivity, which are core principles of Nvidia's organizational culture and management philosophy. Here's the full transcript of the conversation:
When it comes to AI, you have a front-row seat to drive those innovations. What do you see happening in the next five years? These models are getting bigger, the models are getting more concrete. What's going to happen to architecture, data centers, and what's your vision for the next five years?
Probably the most condensed example of computing transformation. As you know, the fundamental change of the computing platform, the transition, it's the basis of the cycle, it's the foundation of every industry that relies on cyclicality, it affects every industry.
What's happening, if you look at your keynote, once you start using accelerated computing, the next thing that might be is that generative AI is added on top of that, and without the transition to accelerated computing, generative AI is going to be very difficult to achieve.
The good thing about moving to accelerated computing is that all of a sudden, something that used to be hard to scale with CPU scaling, and suddenly you're talking about a 1000x X-factor, and there's another 30x X-factor on top of that. On top of that, when you add generative AI, there's another factor of 100,000.
You said at the beginning that the design tool does a single processing, but what the designer wants to do is explore a multi-dimensional, multi-modal, usually exploring space. There is no right answer. There is only one best answer. So we need to explore hundreds of thousands of different areas.
But of course, the exhaustive exploration of the universe, design, is too difficult. Infinite computation can't do it either. So we need AI to help us jump into a specific area of exploration and optimization, and then use the master requirement solver to really focus there. So we can do a lot of different things together. So, accelerated computing is going to change the way Cadence develops software in the first place, and it's going to change the way we use software, that's the first point.
In addition to being able to do it well, there are several benefits. We design our circuits, our chips, our PCBs, our systems, and now even our data centers, and in Cadence, you know that very well. We use you for circuit design, logic design, system design, simulation, verification, formal verification, and more, end-to-end, all the way to the fluid. So the design space now is no longer about a chip or a system, but really a co-design that spans the whole thing.
Millennium is really a great example because you're essentially a co-design company. The way the computer science industry, the computer industry talks about it, the way we talk about it in engineering is co-design, and the way we talk about it in computing is called full stack, but it's the same idea, you have to innovate on the whole thing. So your transformation, from chip design, EDA company to EDA SDA company, is very far-sighted and very necessary. That's exactly how we worked with Cadence, and that's how we designed the system.
Some of the areas that people are starting to realize, your keynote is actually very good, and I would recommend everybody to watch it a few more times because it's very dense, it's very dense.
Either way, one of the areas that you mentioned is very profound, and that is that by investing in accelerated computing, investing in AI, investing in data centers, we can design better, more energy-efficient products.
Now remember, you design chips only once, but you're going to ship trillions of times. You build a data center once, but you save electricity, the 6% power savings that you show in media technology, and that 6% is going to be enjoyed by a billion people all day. So, by designing better software, better ships, better systems, we will be able to save energy for the world, which will have permanent benefits for society.
On the one hand, we're going to consume more power for AI and power for data centers. On the other hand, for 98% of consumption and energy consumption, we're going to reduce that, design better products, design better computers, better cars, better phones, and so on, better materials, and so on.
We're at a very, what people call an inflection point, a transition. All of this is absolutely true. It's a very exciting time, and your keynote really highlights that.
You talked about this transformation, not just making chips, because people would be confused, and of course, Nvidia is making the best chips. Even your key vocabulary on GTC, is building the whole data center, the very complete system, the whole software stack, putting it all together. You have racks and liquid cooling data centers. So the whole architecture, this is not a simple transformation of a typical chip company.
India has turned into a complete software systems company, so I'm curious, how do you do that, or do you always have this idea? or how it unfolds, it's very difficult to implement.
Now some companies are trying, system companies are trying to make chips, it's difficult. But Nvidia has executed that transformation perfectly, from more than just chips to systems to software to data, so I'm curious how you got started with something like this.
I've been a chip designer for a long time. It's been my entire career. There are some things that we observed a long time ago and turned out to be true.
The first thing to do is to observe that a small part of the code in the program represents the vast majority of its runtime. For example, CFD, it may be that 3% of the code represents 99.9999% of the running time. If so, why would you use the exact same tools, the same instruments, the same processors to process all 90%, all 100% of the code? Why don't you do something for 97% of the code and then do something special for 3% of the code, and by doing that, you can speed up the app by 100,000 times. This is observed. Of course, there are very few types of apps that need this benefit and are willing to rewrite.
We wisely chose computer graphics as our first accelerated computing choice, as it turned out to be a very demanding app for a lot of parallel computing. It's good for parallel processing, and it's also a very big market, the market changes very quickly, and innovation is very fast. So we chose a good market as a starting point. But we always imagine that there are a whole bunch of other apps beyond computer graphics.
Accelerated computing is not like general-purpose computing. In general-purpose computing, you can create a processor that can run all the code. This is definitely not the case in accelerated computing, because I coined the term "accelerated computing". What I'm saying is, you accelerate an app, it's an app acceleration computing platform, you have to know what the app is. So in the case of Nvidia, we started with computer graphics, but we also did imaging, and then we did molecular dynamics. I'm excited to see your work on digital biology.
By the way, like a quick one, today we call it drug discovery. We call it drug discovery. Can you imagine if the chip design industry was called the chip discovery industry?
The reason for this is because the engineering team came in and had a chance to see what we found out about Blackwell. And then next year was like a drought, and what happened this year? We didn't find it, and that's perfectly reasonable. But we're never going to do that. Isn't it?
The reason is that biology is much more complex. We shape a transistor until we can use the design tools. You can't shape biology until you can use a design tool. You need design tools to catch up with biology, and that's close.
One of the largest industries in the world is going to turn into a Cadence industry, not the 9% you mentioned, and there's a huge opportunity ahead. Every industry, you mentioned biology, the transportation industry. All apps are different. Some have to do with imaging, some have to do with particle physics, some have to do with fluids, some have to do with finite elements that you know. You write grids and stuff like that. So the algorithm is different.
Cadence is a math and computing company, and in many ways, Nvidia is also a math and computing company. Exactly. That's why we get along so well. And, when Andrew and I got together, what we said to each other, we quickly understood.
When you see the two of us, we're in a restaurant, we're talking about all these things to come. One of us is like this, and the other is like that. That's because we agree with each other. Once I'm like this and he's like that, then we're in trouble.
In any case, we are always about the acceleration of a specific area. Over the course of 30 years, we've accumulated all these different DSLs, domain-specific libraries, some per-particle, some imaging, some AI, and so on, on an architecture called CUDA.
You gave a great presentation at GTC, which was pretty remarkable. As I told Jensen, he'll need a bigger stadium next time. That huge stadium didn't have enough space. So next time you might be in Vegas or wherever you're going. But you've highlighted so many apps that have the level to enable them in almost every industry. But then there are some industries that can have a big impact, like you mentioned life sciences, but you also talked about robotics, or autonomous driving. So is there one or two that you are most excited about in the short or medium term and have the most potential impact?
A couple of industries you mentioned happen to be three that I'm really excited about. One of them is the data center or just computing. The second is, you mentioned cars, but if I could abstract cars as robots, autonomous machines. Autonomous systems or semi-autonomous systems, that big category, whether it's a car or a truck, or a pizza delivery robot or a humanoid robot, a self-propelled robot, these systems have a lot in common. They all need to have multiple sensors, and it is important to require functional safety.
The way you design the computer and validate the computer, validate the computer, certify the computer is very important, the operating system is not your normal operating system. These operating systems are our fail-safe operating systems. So the way you design the system is very detailed and specific.
Of course, the use of AI is very expensive. These systems will always be connected to the cloud, connected to the data center, so that it can, of course, update new experiences, report failures and new situations, and then download new models.
I love the whole autonomous systems space, and a whole new class of devices that all of us are going to build in the near future is going to be humanoid robots, and you can foresee that humanoid robots might be much cheaper to manufacture than people expect. Why is it more than $1~$20,000? You buy a car for $1~$20,000. Why can't you have a human or a robot for $1~$20,000? Robots are likely to be more flexible and versatile than humans in an environment designed for us, in a world designed for humans. So, assembly lines are designed for humans, warehouses are designed for humans, and a lot of things are designed for humans, so humanoid robots are likely to be more productive in that environment.
I love that we're going to turn biology into an engineering field. The scientific discovery process is very important, but it is sporadic. That's why Urm's law is correct. By the way, if we don't move to accelerated computing, if we don't go to AI, the computer industry will experience Om's Law.
The reason is very clear. The amount of work we do, the amount of computation we do is growing all the time. But CPU scaling has slowed down, so we'll enjoy inflation in compute costs rather than reduction. As a result, we have to turn to accelerated computing to save power, save time, and save money.
In any case, digital biology will experience a full-scale renaissance, with science and engineering getting closer and closer. This is a very complex field. Obviously, we don't talk about the Schrödinger equation in chip design because we change the transistor until we can avoid the Schrödinger equation.
Unfortunately, in biology, where covalent bonds, the way chemistry works, obviously the Schrödinger equation is necessary. So we have a lot of things that need to be innovated. But for the first time, we have the necessary tools, computing systems, algorithms, to help us deal with very large and very chaotic systems, the fusion of a data-driven approach with the principled, principled simulation approach that you mentioned earlier, and that convergence might give us a chance.
Because in any case, the car will be like self-driving, probably the first robotic, but then the humanoid robot is another. Lately in biology and in reality, in data centers, so there's a big problem there, which is electricity use.
Cadence designed our data center, and now one computer is the entire data center.
What do you think about the power that AI consumes, the power that data centers consume? Of course we can optimize, but what else can we do? One of those ways is if you do more accelerated computing. In fact, people don't realize that your electricity utilization is definitely going to drop. What are your thoughts on AI and power consumption in data centers?
The first thing you just said is absolutely true, the power usage of accelerated computing is very high. The reason is because computers are very dense, so the power consumption is high. Any optimizations we make to power utilization translate directly into more performance.
This performance can be measured as, more productivity, more revenue generated, or directly translates into savings for the same performance. You can buy something smaller, okay? Power management in accelerated computing translates directly into everything you care about. I've just made some observations that seem to be correct.
Actually, the second thing you just said would imply that it's actually completely wrong. And what we mean, and whatever the root causes are, accelerated computing, as you can see in the demo, with tens of thousands of general-purpose servers, they consume 10~20 times the energy, and the cost is 20~30 times, reduced to a very dense thing.
Therefore, the density of accelerated computing is the reason why people think it is power-hungry and costly. But if you look at it in terms of each ISO job done or ISO throughput, actually, you're saving a lot of money. That's why CPU scaling has slowed down, and we have to move to accelerated computing, because you can't continue to scale in the traditional way. Therefore, accelerated computing is necessary.
The second thing that you emphasized in your keynote is that you have to pay very careful attention. AI actually helps people save energy, and without the AI models that you created that we're using in our tools right now, how would we find 6% more savings or 10x more savings that would not have been possible without AI?
You invest in the training of a model once, and then millions of engineers like us can benefit, billions of people will be able to enjoy savings for decades, and that's the way to think about costs, the way to think about investment, not just on an example basis, but in healthcare, you have to look at the savings vertically, look at climate change, look at the money savings, look at the energy savings, look at the whole span vertically, not just the product that you're building, but the way you're designing the product, and the impact of the use of the product. When you look at it vertically like that, AI will be very helpful in helping us fight climate change, use less electricity, be more energy efficient, and so on.
Jensen, you also have a very unique management style, and your leadership style is well known. Flying organization, quick decision-making, it's incredible to see. And there are a lot of audiences, engineers and managers and leaders. So what do you suggest for these things? How to transform, because to turn ideas into actions.
At the heart of the video management system and leadership philosophy is, if you will, to create the conditions so that amazing people can do the work of their lives. I'm just describing a philosophy, a mission statement, if you will, and how myself and the leaders behave. So the question is, what can we do to create the conditions where people can do the work of their lives?
One of the most important aspects is to empower them with information. So I don't think any decision I make only needs to be heard by one person, or that I need to whisper some information to someone on a one-on-one basis because other people don't deserve to hear it or can't hear it, or, too hard to hear or whatever it is. So I tend to do most of my work in a large environment where there's a diverse team of experts and contributors coming together and we're just solving problems, we're just solving problems. So that's the first point.
I also love the opportunity to reason about things in front of people, in addition to the challenges of the company and the transparency of the information that people should be able to get. This forces us to make recommendations, which are oriented in a direction that is basically based on good reasoning. So by forcing myself to reason about things, I did two things. Of course, I'm influencing others. Second, teach others how to reason, and the ability to reason a very complex, abstract idea to what exactly we should do.
Why we should do this now, or shouldn't do it now, most of the time it's about not doing it now. We don't do it at all. This reasoning process is very empowering for people, so that's one of the reasons why Nvidia is so small.
There were only 28,000 of us, but we hit far beyond our weight because almost everyone was empowered to make good, reasoning-based, principled, first-principles decisions on my behalf. So that's why we do what we do.
Finally, your organizational attributes should reflect the products you make. Nvidia is a full-stack company, we're complete, we're just like you, we're completely co-designed.
Co-design means you shouldn't just do something with the hardware team. You shouldn't just do something with the software team. You should do something on everything at the same time because you are doing it. So I'm trying to create an environment where experts and contributors at every level of the company can be involved in solving problems at the same time, and at the same time participating in problems, those are some of the principles.
PS: Looking back at 2023 so far, most of the AI hype has focused on the horizontal capabilities of the underlying model, but the real opportunity for AI lies in AI and how Agent reconfigures and creates the B2B value chain, with 112 top VCs selecting the top 30 tech startups in 2024, nearly 50% of which are GenAI, and less than 1/4 of SaaS.
Reference：
Hatps://vv.youtube.com/watch?v=yurlak2kugm
Author: Youxin;Source: Youxin Newin
This article was originally published by @有新Newin on Everyone is a Product Manager and is not allowed to be reproduced without permission
The title image is from Unsplash and is licensed under CC0
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.
![China has become a popular market for life sciences investment in the Asia-Pacific region in 2023[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)