Nvidia CTO Michael Kagan：突破摩尔定律，迈向百万 GPU 集群

One of the interesting things about NVIDIA is the culture of win-win. We are not after taking a bigger piece of the existing pie. We are after baking a bigger pie for everybody. Our success is our customer's success. Our success is not the failure of our competition. And I think fusing together conventional computing—von Neumann machines—and accelerated computing provided by NVIDIA actually gives NVIDIA and Intel channels to the market. We're expanding the market and serving the markets that otherwise were more challenging.

Yeah.

We're delighted to hear today from one of the legends of the semiconductor industry, Michael Kagan, the CTO of NVIDIA. Michael was formerly chief architect at Intel, and then co-founder and CTO of Mellanox, which NVIDIA acquired for $7 billion in March 2019.

In the time since, Michael has been a major driver of NVIDIA's dominance as the AI compute platform, in large part due to the role of Mellanox and interconnect in driving chips beyond Moore's law. The AI race is ultimately a silicon race to squeeze the most intelligence possible out of each unit of silicon, and Michael takes us on a journey through how the compute frontier has evolved—from squeezing more transistors onto a single chip to bringing together thousands and hundreds of thousands of chips into a single fabric connected by networking in an AI data center.

Michael has been driving the compute frontier forward for more than 4 decades, and we're honored to have him on today's show.

Okay. We're here with Michael Kagan, the CTO of NVIDIA, currently the world's most valuable company. Michael, thank you for joining us.

Thank you. My pleasure.

I thought we could start with this: Our partner Sean likes to make the case about every 6 months that NVIDIA would not be NVIDIA without Mellanox. Mellanox is a company that you co-founded some 25 years ago and have been a part of through this day. Can you paint that picture for us? Why was the Mellanox acquisition so critical to NVIDIA?

There is a huge transition in the world in terms of computing and the need for computing. It grows exponentially, while one of the things that we usually estimate linearly is actually exponential in the world. And exponential growth is now actually accelerating.

It used to be like Moore's law, which was based on basic silicon: 2× every other year. Regardless of the discussion that Moore's law, in terms of physics, is not quite running anymore, once AI kicked in—which was in 2010 or 2011—it kicked in when a GPU, from a graphics processing unit, became a general-purpose computing unit.

Yeah.

That was when workloads were run—the first time an AI workload was run on the GPU—taking advantage of the programmability and parallel nature of this machine. The requirements for performance started to grow at a much higher rate. The models started to grow in terms of size and capacity, 2× every 3 months, which now requires 10× or 16× a year in performance growth versus the old school of 2× every other year.

In order to grow at this scale, you need to innovate and develop solutions at a much higher scale than just a basic component. That's where networking comes in.

Mm-hmm.

There are multiple layers of scaling performance that require high-speed and high-performance networks. One is what we call scale-up.

Mm-hmm.

Basically, going back to the CPU days, scaling up was Moore's law—more transistors—and also advances in the microarchitecture, like out-of-order execution and, at some point, multicore, and so on and so forth. This is the basic building block of computing. In the GPU world, the basic building block is the GPU.

Mm-hmm.

In order to scale it up more than you can on a single piece of silicon, with all the advances that we are making in microarchitecture and advanced technologies, you actually need to do something on the scale of a multicore CPU, but at a much larger scale. That's what we are doing with NVLink. That's the scale-up solution.

Our GPU—what we call a GPU today—is a rack-size machine.

Yeah.

You need a forklift to lift it.

Yeah.

If you order just a GPU on Amazon, don't be surprised if this huge rack shows up.

Yeah, people think chip, but it's really a system.

Right. Right. And that's just one GPU.

Yeah.

So, the basic building block—a very basic computer on which application software is running—is this GPU. And it is not just silicon, not just hardware, and not just wires; there is also a software layer that exposes CUDA as the API.

That's what enables you to pretty much seamlessly scale. I'm simplifying the story a little bit, but you can seamlessly scale from one component, which used to be a single GPU, all the way up to 72, maintaining the same software interface.

Once you get this building block as big as it conceivably can be built—in terms of power, cost, and efficiency—then you start scaling out.

Yeah.

Scale-out means you take many of these building blocks, connect them together, and now, at the algorithm level and the application level, you actually split your application into multiple pieces running in parallel on these big machines.

Mm-hmm.

And that's again where networking comes in. If you talk about scale-up, we basically made a memory-like domain to go beyond a single compute node or a single GPU.

Mm-hmm.

That's actually the first place where Mellanox technology comes in. Before the Mellanox acquisition, the scaling-up of NVIDIA with NVLink was limited to a single-node machine.

Going outside of a single compute node—those 72 GPUs are actually 36 computers, each with 2 GPUs wired together to present all of this as a single GPU—is not just plugging a wire into the connector. There is a lot of software and a lot of technology within the network involved in making multiple nodes work as a single machine.

That's where Mellanox first came in. In terms of the way we go up in scale, that's the first one. The second one is: How do you split the operation across multiple machines?

Yeah.

The way to do it is this: If I have a task that takes 1 GPU 1 second to do, and I want to accelerate it, I split it into 10 or 1,000 pieces and send each piece to a different GPU. Now, in 1 millisecond, I get done whatever I was doing in a second.

But you need to communicate this partial job split. You split the task, and then you need to consolidate the results. Every time you run this, you have multiple iterations or multiple applications running at once. There is a part of doing communication and a part of doing computation.

Mm-hmm.

The thing is that you want to split it into as many pieces as you possibly can because that's your speedup factor. But if your communication is blocking you, you waste time, you waste energy—you waste everything.

What you need is very fast communication. You split it into many, many pieces, so each piece takes very little time, but then there is another piece that is communicated, and you need to feed it in time. That's just pure bandwidth.

Another thing is that when you tune your application, you tune it so that communication can be hidden behind computation. That means if communication, for some reason, gets longer, then everybody waits.

Hmm.

It means that what you need in the network is not only raw performance—what are called hero numbers: “I can get to that many gigabits per second.” You also need to make sure that, no matter who communicates with whom, the latency—the time it takes—is distributed very narrowly.

Mm-hmm.

If you look at other network technologies or other network products, you go to the hero numbers: sending a bit from one place to another. That's basically physics.

Yeah.

It's pretty much close to the same for everyone. We are a little bit better, but that's not the big advantage.

When you do it thousands of times, and it takes the same amount of time to do it every time versus having a very wide distribution with other technologies, the machine becomes less efficient. Instead of being able to split your job across 1,000 GPUs, you can split it across only 10 GPUs because you need to accommodate for the jitter on the network within the computation phase.

Inherently, the network determines the performance of this cluster. We look at this data center as basically a single unit of computing.

Yeah.

A single unit of computing means that you start architecting your components, your software, and your hardware at the point where this is a data center. This is 100,000 GPUs that we want to make work together. We need multiple chips: 2 compute chips and 5 network chips.

This gives you a sense of the scale, the impact, and the investment you need to make to create this single unit of computing. That's where Mellanox technology came in.

Another aspect of this is that we talked about the network that connects the GPUs to run the task. But there is another side of this machine that is customer-facing. This machine needs to serve multiple tenants, and it needs to run an operating system. Every computer runs an operating system.

Another part of Mellanox technology is what we call the BlueField DPU, or data processing unit, which is actually the computing platform used to run the operating system of the data center.

Mm.

In a conventional computer, we have a CPU that runs the operating system and application software. There are many things we can talk about in terms of the advantages and disadvantages, but there are 2 key things. One is how much time you spend on general-purpose computing to run the application. You want to maximize it.

Mm.

Another thing is how you isolate your infrastructure computing from your application computing.

Mm.

Because of viruses, cyberattacks, and so on, being able to run infrastructure computing on a different computing platform significantly reduces the attack surface, especially for side-channel attacks.

Mm.

This is different from what happens if you run it on the same computer. If you remember, about 5 or 6—well, actually, almost 10—years ago, there was Meltdown and all these side-channel cyberattacks on CPUs. This cannot happen, or the attack surface is significantly reduced, when you run them on different platforms.

On the other side of the network, we also have technology. That's what makes the data center more efficient. I may not be objective, but I do agree that the merger of Mellanox and NVIDIA actually goes both ways.

I don't think the networking business—now NVIDIA, previously Mellanox—could have grown that significantly at the rate it has grown. Now, I think we are the fastest-growing networking business, let alone NVLink and InfiniBand. But just the Ethernet business is the fastest-growing business ever.

Yeah.

What are the things that break as you get to 100,000, and maybe eventually 1 million, GPU clusters? How do you use software to help design around that?

It's a multistage challenge.

One of the things you need to keep in mind, which is not very obvious to all engineers when you design a machine or think about how to operate it, is that you have these components, they're working, and now you just need to figure it out.

The hardware component works at 99.999-whatever percent of the time. That's usually okay if you're dealing with a single box or a couple of them. But if you're building a 100,000-component machine—a 100,000-GPU machine—which means there are millions of components, the chance that everything works is zero. Something is definitely broken.

You need to design it from both a hardware and a software perspective to keep going as efficiently as you can, maintain performance, keep your power efficiency, and, of course, keep the service running. This is challenge number 1, even before we get to millions. This challenge actually starts at a few tens of thousands.

Number 2 is that, when you're running these workloads, it is really important to write the software and provide all the interfaces needed to place the different parts of the job more efficiently. Sometimes you run a single job on the entire data center, and then you need to place the different parts of the job efficiently.

Building networks at this scale is a very different story. Building a compute network at this scale is very different from building a general-purpose data-center network.

Yeah.

A general-purpose data-center network is the Internet. It's not a big deal. Well, it is a big deal, but it's a different deal. You're serving loosely coupled, collaborative microservices that create the service you see as a customer from the outside. Here, you're running one single application on 100,000 machines.

Yeah. Is that specific to training workloads, or is that also true for inference workloads?

It's true for everything. It depends on the scale. Inference is yet another topic that we may touch on.

Until recently, training was the key thing. There were a lot of GPUs, and training was being done in a very specific way. You basically copy the same model onto multiple machines or multiple sets of machines and run them, then consolidate the results, and so on.

With inference, the story is a little different. But you need to provide the hooks in the hardware and in your low-level system software for applications and schedulers to place the job and the different parts of the job in the most efficient way.

As long as your machine fits in a building, which is about 100,000 GPUs, that's one thing. Now, when you're talking about gigawatts, it's all power-driven. The challenge is that, for many reasons, you want to split your workloads across multiple data centers.

Sometimes data centers are many kilometers, many miles apart. They may be across the continent. This comes with yet another challenge, which is the speed of light.

Yeah.

The latency variance between different parts of your machine is dramatically different. What's even more challenging is that, when you talk about networks, congestion on the network is one of the key problems that deteriorates network performance.

Yeah.

Managing congestion across such a latency difference is not like in the old telecom days, when you put a box at the edge of your data center with huge buffers and it acts as a shock absorber for congestion. A huge buffer is not good. Bigger is not better. There is a famous statement from a very famous one.

These buffers, or these devices, are basically there to isolate the external world from the internals. But when you want to run a single workload across data centers that are separated by kilometers, every machine on one side needs to be aware of whom it communicates with, whether it's a short communication or a long communication, and adjust all the communication patterns accordingly.

You don't need these big buffers, because a big buffer creates jitter.

Yeah.

Mm-hmm.

We have a technology that we actually developed recently. All of our Ethernet networking is Spectrum-X. This is a technology that we designed and developed based on the Spectrum switch that we put on the edge of the data center, and it provides all the information and telemetry needed for the endpoints to adjust for congestion.

Yeah.

Can we talk a little bit more about training versus inference? How does the shape of the workload differ?

Mm-hmm.

I guess backpropagation is a lot more computationally intensive, and the forward pass is less so. But how does the workload differ?

Mm.

Are you seeing customer demand start to shift from pretraining toward inference, or do you think it's still very training-heavy right now?

And if I could just ask a quick follow-up question: Will people be running inference workloads on the same data centers that they use for training, or will these end up being 2 separate systems? Because they're different optimizations, will people end up using 2 different sets of data centers?

Okay, yeah. That's a great question. Let me start with the first one.

Training has 2 phases. One is inference, which is just forward propagation, and then backpropagation to adjust the weights. For data-parallel training, there is yet another phase to consolidate the results of the weight updates across multiple model copies.

Until recently, training was the main driver of compute, because until not very long ago—maybe 2 years, which is ages in the AI era—inference in AI was mainly perceptual.

Mm.

You show the picture: That’s a dog. You show the photo of the person, and here’s Michael and here’s Sonya, and that’s it. That’s a single path, and that’s it. Then came generative AI, where you actually get recursive generation. When you pose a prompt, it’s not just 1 inference.

Mm-hmm.

It’s many inferences. For every token, when you generate text or a picture, you need to go through the entire machine all over again. So instead of one-shot inference, there’s more. And now there’s reasoning, which means the machine starts thinking.

Yeah.

If you ask me what time it is now, I can tell you. It’s easy, right? But if you ask me a more complicated question, then I need to think. I probably need to wait or compare multiple solutions or multiple paths. Every such thing is inference.

Mm.

Every such thing is inference. Inference itself has 2 phases. One is much more compute-intensive, and the other one is memory-intensive.

Mm.

It’s what we call prefill, because when you do inference, you have some sort of background, right? That’s the prompt—some relevant data that you need to process and create the context to generate the answer. This is very compute-intensive; it’s not very memory-intensive. The other part is actually generating the answer, which is the decode part of inference, where you generate token by token.

Mm-hmm.

There are some technologies that let you generate more than 1 token, but it’s still a single path, much less than the final answer.

Mm.

If you combine all these things together, inference demand for computing is actually not less than training.

Mm.

It’s actually even more. There are 2 reasons for this. One is what I explained: There’s much more computing than there used to be for inference. The other thing is, you train a model once, but you infer many times.

Yeah.

ChatGPT has almost 1 billion people, right? There are customers; they’re pounding it all the time with the same model. They trained it once.

More than 1 billion.

Right. Now they’re making videos, so that’s a lot of inference. You can generate, and everybody is doing the inference. My wife, I think, talks to ChatGPT more than she talks to me these days. Once she discovered it, it became her best friend.

In terms of inference, to your question about machines, you can do inference on the phone. Okay? So there are definitely going to be much smaller-scale installations for inference.

Hmm.

It’s like mobile devices. If you look at data-center scale, the efficiency of programming and programmability is much more valuable than hardware optimization. Every hardware instance has its own cost and its own drawback. It’s a very similar GPU, with the same programming model as a GPU for prefill versus decode. I don’t remember when it happened, but we announced that we are building a GPU SKU optimized for prefill.

Hmm.

Hmm.

You will have a GPU that can do decode, and a decode GPU can do prefill. You can equip your data center with SKUs for prefill versus SKUs for decode to optimize for typical use.

Yeah.

But if your workload shifts toward more decode or more prefill, you can use either one of them to compensate. This is the importance of programmability: the same interfaces for GPUs. It’s based on CUDA and UP, which is what made NVIDIA NVIDIA before Mellanox. Yeah, yeah.

Can I ask you a question about data-center scaling? For many decades, we had Moore’s Law, and chips got more and more dense and produced better and better performance. Then we ran into the laws of physics.

Chips just couldn’t get more dense because their quantum-mechanical properties caused them to break down. So then we had to scale up to the rack level, and now we have to scale out to the data-center level. Is there some analogous law of data-center scaling that says when data centers get too big, the communication overhead causes the performance to break down? Or, said differently, is there a natural limit to how big data centers can get?

I think there is a practical limit to how much energy you can consume within a given size of data center.

But if you were surrounded by nuclear power plants and the energy was available, would the—

Well, the—

…would the data center itself perform?

I don’t know. I’m not an expert in construction, even. But if you surround it with nuclear power plants, there’s energy coming in. Now the heat is going out.

Yeah.

There’s a whole bunch of technology coming to help make this more and more dense. We have now moved pretty much entirely to liquid cooling.

Yeah.

One of the reasons we did it is to enable much denser compute power. We couldn’t build computing as dense as what we’re building now with air cooling. The last big data center, which is, like, xAI-scale, is 100 or 150 megawatts now. We’re talking about gigawatt data centers. People are talking about 10-gigawatt data centers. There’s a desire to build much, much bigger data centers.

Are you sending the data centers to outer space?

Free cooling, free power.

I think one of the things that determines the speed of data-center deployment is how fast concrete gets stable.

Fair enough.

Before starting Mellanox, you were at Intel for 16 years?

That’s right.

Sixteen years. You became chief architect? NVIDIA and Intel recently announced a partnership. Can you share a little bit about what the vision for that might be?

The starting point is that computing changed in the last decade, or a little bit more than a decade. NVIDIA started as the accelerated-computing company. Video games were the first application, and then it evolved to AI, which is the new way of data processing.

A general von Neumann machine just isn’t capable of being used as a platform to solve the problem. Programming a von Neumann machine is just explaining to somebody what to do. I can explain many things, and I can explain to many people what to do, but I can’t explain how to distinguish between a cat and a dog, right? There are new challenges that AI solves, and you need acceleration there.

Our partnership with Intel is actually fusing accelerated computing with general-purpose computing.

Mm-hmm.

General-purpose computing isn’t going away. Everything will be accelerated, but we accelerate general-purpose computing. We accelerate the applications. x86 is the architecture that is dominant there, and it would serve both companies greatly.

That’s one of the interesting things about NVIDIA: It’s the culture of win-win. We’re not after taking a bigger piece of the existing pie. We’re after baking a bigger pie for everybody. Our success is our customer success. Our success is not the failure of our competition. Our success is the success of our customers and the success of our ecosystem.

I think fusing together conventional computing, von Neumann machines, and accelerated computing provided by NVIDIA probably opens yet another dimension that I’m not sure what it is, but in the practical, short-term view, this gives NVIDIA and Intel channels to the market, expanding the market and serving markets that otherwise were more challenging.

Mm.

You mentioned the culture of NVIDIA. When Mellanox became part of NVIDIA in 2019, the market cap of the combined company was about $100 billion, which is no joke. But the market cap today is about $4.5 trillion.

That's right.

A 45× growth in value in 6 years is pretty phenomenal. How has that changed the culture of NVIDIA? How is NVIDIA different today, now that it's one of the most admired companies in the world, if not the most admired, versus 6 years ago?

When we just joined, Jensen was in Israel, and I presented him with my belief that one plus one would be 10. I was actually off by a factor of 4. But Mellanox and NVIDIA, in a sense, are similar. The cultures were very similar to begin with, but nothing is absolutely similar.

I was the only founder who left Mellanox after Eyal resigned a few months after the acquisition. My main focus at the beginning—the things you think about in the shower—was how to make sure that this acquisition would succeed.

Yeah.

NVIDIA paid $7 billion for a company that I founded. With all the mixed feelings that were there, once it's done, it's done. Now I have to make it successful.

Yeah.

Eventually, it worked. Most of the Israeli employees stayed. I think it was 85% or 90% of the original employees.

Wow.

Actually, NVIDIA grew more than 2× in Israel in terms of manpower.

Yeah.

We're growing, and we're announcing that we're actually going to build a new NVIDIA campus in Israel.

Nice.

That's where I think the overall merger was very successful. I did my best to make sure it succeeded. Besides the technology that I was looking at—which is sort of technology, but it's technology and theology—there were many other things to make sure that people were comfortable, so that being at the center of Mellanox, whose headquarters were in Israel, they didn't feel left somewhere far away.

Jensen basically emphasizes networking as the critical part of NVIDIA's success.

Yeah.

And he's right.

Yeah, yeah, yeah.

I think it's considered to be the most successful merger in the history of technology. You guys probably track these things better than I am, but overall, I think it was a great move.

Yeah.

What are the science-fiction things that you spend your time thinking about, just wondering about? For example, optical interconnects: Do you think those will exist? Do you think AI will ever be better at physics than us, and better at data-center design than us? What do you think about?

What I'm thinking, if you look at science fiction, is how to make history experimental science.

Mm.

In physics, you try something, see if it works, and then try something else. In history, time goes in one direction. But if you have a good simulation of the world, you can make history experimental. We have Earth-2, a climate simulator.

Mm.

With this type of technology, we can actually simulate how what we do today will impact global warming 50 years from now.

Mm.

Mm.

So, experimental science. You try something and see what happens 50 years later. That's the science-fiction part.

Yeah.

On physics, now we are moving from reasoning and so on and so forth. Once we get AI models to understand physics, we can actually learn physics.

Yeah.

AI can teach us physics because the way we get to the laws of physics that we observe— theoretical physics, right?—is that you observe some phenomena, generalize it, and compose the rule, basically the physics law, that stays underneath these phenomena. AI is really great at generalizing, data processing, and observing. So AI can help us get to know some laws of physics that we don't even imagine now.

Yeah.

Moore's Law was 2× every 2 years. Huang-plus-Kagan's Law is... What is the slope, and how long do you think you can sustain it?

The slope is somewhere in the range of 10×, or a few orders of magnitude, a year. That's what we are doing now, by the way. About 2 or 3 years ago, we accelerated our product introduction from every other year to every year.

Now we introduce a new wave of products every year, and it's an order of magnitude higher performance. It's not chip-level performance. It's the machine that you can build with this performance. That's what we are looking at: a single unit of computing.

How long it will stay, I don't know. We'll do our best to maintain it as long as needed and probably even accelerate. It's all about the exponent. It's hard to imagine.

If you look at Moore's Law curves, or any law curves, they usually plot them on a logarithmic scale. So it looks linear, but that's the wrong thing to look at. You can't predict what's going to happen.

Who could predict that when the iPhone was first introduced, or when the smartphone was first introduced—that's 15 years ago?

2007 was the iPhone.

'07.

Yeah, 2007. Oh, 17 years ago. Who could imagine that the least-used function of this smartphone, at least for me, is a phone?

Yeah.

All of this is e-commerce, texting, news, and mail. It's basically running your life from this machine.

Yeah.

It's your authentication; your ID is there. Who can imagine what's going to happen 10 years from now with all these developments that we are doing today? But we are building the platform for innovation.

Notwithstanding your commentary on “who can imagine,” what is the most optimistic view of our future with AI that you like to think about? What could AI do for the world 5, 10, or 15 years from now?

Steve Jobs called the computer the bicycle of the mind.

Yeah.

AI is maybe—I don't know if it's... It's probably the spaceship of the mind. There are a lot of things that I would like to do, but I just don't have enough time or resources to do them. With AI, I will have that.

It doesn't mean that I will do twice as much. Maybe I will do 10 times as much. But I will want to do 100 times as much as I want to do today.

You go to any project leader, and nobody says, “I have enough. I have enough manpower. I have enough resources. I don't need any more.” If you give him a resource that is twice as efficient, he'll do 4 times more. And he'll want to do 10 times more.

It's like electricity changing the world, right? In London, you still see these gas lamps and the infrastructure to use gas as the source of energy. Who could think that once electricity was invented, it would change the world, and that we couldn't live without electricity?

Mm-hmm.

The same with AI.

Beautifully said. Thank you so much for joining us today. I love this conversation.

Thank you.

Thank you.

Thank you for having me.