NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Well, good morning, good afternoon, everyone. Thank you for joining us. My name is C.J. Muse with Evercore ISI. And it is my distinct honor to host the NVIDIA team here today. We have Paresh Kharya, Senior Director of Product Management in the Data Center; and also on the line is Simona Jankowski, VP, Head of Investor Relations and Strategic Finance. So Paresh, thank you for joining us.

Thank you for having me, C.J. My pleasure.

Excellent. For the audience, there is a chat box where you can ask questions, and I'll do my best to integrate those questions into our fireside chat.

But Paresh, you've been busy. New product cycle, so I figured we would just start there with the new A100 offering. Ampere appears to be a transformative architecture and technology. Can you walk us through what this means for customers in terms of their ability to harness the power of AI?

Yes. So C.J., if you look at, there are 2 megatrends in AI. The first trend is the exponential growing complexity of AI models. At the launch of A100, when we spoke, we talked about how, since the launch of our Volta, the model complexity has grown 3,000x to train the largest models. In fact, that trend continues. Just a month after we announced Ampere, OpenAI introduced a model called GPT-3. It's a 175 billion parameter model, and that's 10x more compute intensity to train these models. So that continues, that trajectory continues, 30x more compute intensity to train the largest model since we launched Volta. And this requires large-scale computers because otherwise, it would just take years to train these massive models. So that's the first trend. The second trend is because these AI models are now so accurate, there is a pervasive AI-powered applications everywhere, from conversational AI to recommender systems and so on. And they service millions of users. And as they service millions of users, each interaction with a user requires a tiny amount of acceleration. So it requires many small-scale accelerated computers. Now these 2 are seemingly divergent needs, but A100, for the first time, provided a single solution for both of these divergent needs at the same time. First, it provides a massive performance boost, 20x higher performance compared to Volta. Second, it provides 50x higher scalability. So a single server can be used as one big GPU working on large-scale training problems or can be interconnected over the network to do even larger-scale training problems. And secondly, with the capability called multi-instance GPU, a server can be partitioned into 50 different smaller-scale GPUs, each running a different inference application. So for the first time, A100 unifies data center acceleration for both training and inference. And that's why our customers are just loving the A100 architecture, and we've really seen a rapid uptake.

Well, that brings me to the next question, which was, can you share with us the reception so far? And I guess I'd love to hear whether it's scale up, scale out or other drivers as you think about the adoption for hyperscale versus enterprise versus supercompute. What excites each of those segments perhaps more than something else?

Sure, yes. So there's been a tremendous uptake. A100 came to the cloud faster than any NVIDIA GPU in the history. It's already available on Google Cloud, on Microsoft Azure, and it's coming soon on every other hyperscale and other cloud providers. There have been over 50 different A100 server models that were announced by the leading server manufacturers within a few weeks of our launch. And each of these will help the vertical industries ramp up over the next coming quarters. And the value proposition is really strong. So for hyperscale with multi-instance GPU, A100 enables great economics, as we discussed, 50x higher scalability as well as a unified training and inference platform. So they can provision instances of different capability from the same infrastructure. And that's why cloud providers and hyperscalers are so excited to bring A100 to their clouds. For enterprises, this enables unification of acceleration in their data center, so they can optimize with a single architecture while optimizing the utility and utilization because the same investment to accelerate the data center can accelerate a full range of applications, whether it's data analytics, artificial intelligence, AI inferencing, virtual graphics, all of that can be accelerated by the same infrastructure and the same investment. Finally, for supercomputing, you mentioned, we introduced next generation of NVLink. And combined with our Mellanox InfiniBand technology, along with our software stack that we call Magnum IO, we can scale up applications massively, something that supercomputing really enjoys, multi-GPU, multi-node, really large-scale applications, all backed up by the same single architecture that they've come to rely upon from NVIDIA. And we've had several design wins in the supercomputing because of this reason. In U.S., for example, with a supercomputer called Perlmutter from NERSC or in Europe with Jülich Supercomputing, with Max Planck Institute and so on, we had several design wins with A100.

Well, congratulations on that. I was hoping to pivot to your system-level approach, which I think is something that is perhaps still underappreciated for NVIDIA, but I'm thinking less so as every day passes. So obviously, acceleration of computing starts with the GPU, but it continues through system design, system software algorithms and optimized applications. Can you speak to the importance of this platform approach for NVIDIA?

Yes. The platform approach is critical. We make great GPUs. Tensor Cores, now in its third generation, is unrivaled in terms of performance. Our NVLink provides really massive speed up in a multi-GPU environment and so on. But it's the 2 million NVIDIA developers that are rapidly growing that's a testament to our platform approach. So we provide our fully end-to-end platform to our developers; our CUDA, which is a single programming model that sits on top of our GPUs; and CUDA-X, which is our layer of domain-specific libraries that provide the map that's needed for the different domains of applications, whether it's AI training, inferencing, running larger-scale jobs and so on. It's that layer of CUDA-X, which is millions of lines of code. And it has evolved for over 15 years now. That's really important. And AI frameworks, they sit on top of it. So frameworks like TensorFlow and PyTorch, the reason why they're able to perform so well on our GPUs is because of the layer, CUDA-X, which contains all the acceleration libraries that make these frameworks work well. But we are not stopping there. We also have now created application frameworks that are really important for enterprises to deploy and create applications, whether it's conversational AI applications with our framework called Jarvis or recommendation applications with our framework called Merlin and so on. And all of the software stack that we have is delivered through NGC, which is our hub for GPU-optimized software. So the net effect is, with the full-platform approach, we are able to sustain a virtuous cycle of adoption. Because the more developers because they find it easy to develop on our platform, it implies creating more applications and more deployed endpoints. Because there are more deployed endpoints, it attracts even more developers on our platform. So because of our full-platform approach, we are able to sustain this virtuous cycle of adoption that's so vital in this industry.

I guess playing devil's advocate here, talking to some folks in the industry though, clearly, I think this is more 1 to 2 years ago. There's a notion that if you can just support 1 or 2 AI frameworks, that's enough and that CUDA, CUDA-X isn't necessary. So can you speak to why you disagree with that notion?

Yes. So the reason why frameworks like TensorFlow and PyTorch work so well on our platform is because of CUDA and CUDA-X. These frameworks sit on top of our CUDA and CUDA-X, and they automatically benefit from all the capabilities that we build in our hardware. For example, we launched our third generation of Tensor Core GPUs with Ampere architecture, and we introduced a new precision called TF32. There is 0 code change that is required from the frameworks because these frameworks use our libraries like cuDNN and so on. And automatically, the benefit of TF32, the new precision that we introduced, comes to developers. So the CUDA and CUDA-X are really important in accelerating all the frameworks. This is the reason why, frankly, our customers confidently invest in our platform because they've seen our commitment with our libraries and our single CUDA programming stack. They know that with every generation of our GPU, they'll automatically benefit from the architecture, sometimes without changing a single line of code. And that's really the benefit that you get when you have platform in layers with the industry framework sitting on top of things like CUDA and CUDA-X.

Great and helpful explanation. I guess curious, as you think about, again, platform approach, are you seeing any differences in terms of what is required by hyperscalers versus your enterprise customers? And I know that we're early in the enterprise ramp for the A100, but would be curious if you're seeing any sort of differences there.

Yes. So with our platform approach, the great advantage we have is we are able to meet where our customers are. Meaning, on one hand, we create our end-to-end solutions like DGX for enterprises, but at the same time, we open up our platform with HGX, which is the guts behind the DGX that hyperscalers deploy as they create their infrastructures as well as our OEM server maker partners who work with us at that layer. So that's on the hardware front. And if you look at the software stack, with the hyperscalers, our approach is to deeply integrate into their software stacks, into the frameworks that they are promoting. So for example, let's take TensorRT. It's our library for optimizing trained models for inference. It's deeply integrated with TensorFlow. Similarly, it's deeply integrated with Microsoft ONNX Runtime. So with the hyperscalers, our approach is to deeply integrate our software stack. On the other side, hyperscalers also contribute to our software stack. Just last week, for example, Microsoft announced DeepSpeed. It's used for extreme scale training of models on thousands of GPUs for models that reach trillions of parameters, really important contribution by Microsoft. And it benefits. Because it's an open source project, it benefits every NVIDIA customer. So that's our strategy with hyperscalers. For enterprises, on the other hand, they don't always have in-house AI or software capabilities. So our approach is to continue to provide higher-level stacks and value prop. NGC is a critical part of that. With NGC, we provide pretrained models so enterprises don't have to start from scratch with training models and so on. They can just take pretrained models that are state of the art, whether it's for conversational AI or recommenders and so on, and they can apply their own proprietary data to customize it to their needs using what's called transfer learning. We also provide application frameworks, which are again targeted for enterprises which have end-to-end workflows, like how we were describing, whether it's Isaac for robotics or Clara for health care. So while hyperscalers led the adoption of AI and accelerated computing, enterprises are now on that same adoption curve. And for many, it's an existential imperative, frankly. And we've seen tremendous growth in the vertical industry. Our data center business derives close to half of our revenues from enterprises.

That's helpful. So you've made a splash announcement, I guess, a week back with the potential Arm acquisition. Curious, as you think about full-stack solution, end to end, how does that asset fit into your go-to-market strategy? And how excited are you to potentially have CPU for the data center in your portfolio?

Yes. So as you pointed out, first, it creates a premier computing company for the age of AI by combining our AI computing platform with Arm's vast CPU ecosystem. Secondly, it expands Arm's IP licensing portfolio with NVIDIA technology in several large end markets like mobile and PC. And thirdly, it turbocharges Arm's server CPU road map pace, which can then further accelerate the data center and edge AI and IoT opportunities for us together. And finally, it expands the reach of our computing platform from 2 million developers that we have today to 15 million developers on Arm. And if you look at from an endpoint perspective, we sell about 100 million-or-so chips per year. Arm architecture-based chips sell 22 billion per year. So ultimately, the benefit is the virtuous cycle. More endpoints and applications will attract more developers. And more developers will result in even more applications and more deployed endpoints. And so that's the exciting possibility here.

That's great. Appreciate that. Hoping to pivot now to different business verticals. I think something that's special to NVIDIA is that you're clearly taking a customized approach to each vertical. And I guess, can you speak to which businesses are currently the fastest to adopt accelerated computing for AI? And how are they benefiting from your platform?