 From around the globe, it's theCUBE. Covering HPE Discover Virtual Experience, brought to you by HPE. Hi, I'm Stu Miniman and this is the Cube's coverage of HPE Discover, the virtual experience for 2020. Getting to talk to HPE executives, their partners, the ecosystem, where they are around the globe. This session we're going to be digging in about artificial intelligence. Obviously a super important topic these days, and to help me do that, I've got two guests from NVIDIA. Sitting in the window next to me, we have Presh Karya, he's the director of product marketing and sitting next to him in the virtual environment is Kevin Deerling, who is the senior vice president of marketing, as I mentioned, both with NVIDIA. Thank you both so much for joining us. Thank you Stu, great to be here. Great to be here. All right, so Presh, why don't you set the stage for us? AI obviously, one of those mega trends to talk about, but just give us the stages to where NVIDIA sits, where the market fits, and your customers today is that they think about AI. Yeah, yeah, so we are basically witnessing a massive changes that are happening across every industry. And it's basically the confluence of three things. One is of course AI, the second is 5G and IoT, and the third is the ability to process all of the data that we have that's now possible. For AI, we are now seeing really advanced models from computer vision to understanding natural language to the ability to speak in conversational terms. In terms of IoT and 5G, there are billions of devices that are sensing and inferring information. And now we have the ability to act, make decisions in various industries. And finally, all of the processing capabilities that we have today at the data center and in the cloud, as well as at the edge, with the GPUs, as well as advanced networking that's available, we can now make sense of all of this data to help industrial transformation. Yeah, Kevin, you know, it's interesting. When you look at some of these waves of technology and we say, okay, there's a lot of new pieces here. You talk about 5G, it's the next generation, but architecturally, some of these things remind us of the past. So when I look at some of these architectures, I think about what we've done for high-performance computing for a long time. Obviously, you know, Melanox, where you came from through NVIDIA's acquisition, you know, strong play in that environment. So maybe give us a little bit, compare contrast, you know, what's the same or what's different about, you know, this highly distributed, you know, edge compute AI, IoT environment. And what's the same with what we were doing with APC in the past? Yeah, so we've, you know, Melanox has now been a part of NVIDIA for a little over a month. And it's great to be part of that. We were both focused on accelerated computing and high-performance computing. And to do that, what it means is the scale and the type of problems that we're trying to solve are just simply too large to fit into a single computer. So if that's the case, then you connect a lot of computers. And Jensen talked about this recently at the GTC Keynote where he said that the new unit of computing is really the data center. So it's no longer the box that sits on your desk or even in a rack, it's the entire data center because that's the scale of the types of problems that we're solving. And so the notion of scale up and scale out the network becomes really, really critical. And we're doing high-performance networking for a long time. When you move to the edge, instead of having, you know, a single data center with 10,000 computers, you have 10,000 data centers. Each of which has a small number of servers that is processing all of that information that's coming in. But in a sense, the problems are very, very similar. Whether you're at the edge or you're doing massive HPC scientific computing or cloud computing. And so we're excited to be part of bringing together the AI and the networking because they are really optimizing at the data center scale across the entire stack. All right, so it's interesting. You mentioned, you know, NVIDIA's CEO, Jensen. I believe if I saw right in there, he actually said, throw it a term which I'd not run across. It was the data processing unit or DPU in that, you know, data center as you talked about. Help us wrap our heads around this little bit. You know, I know my CPUs, you know, when I think about GPUs, I obviously think of NVIDIA, you know, TPUs in the cloud and everything we're doing. So, you know, what is DPUs? Is this just, you know, some new AI thing or you know, is this kind of a new architectural model? Yeah, I think what Jensen highlighted is that there's three key elements of this accelerated, disaggregated infrastructure that the data center is becoming. And so that's the CPU which is doing traditional single threaded workload. But for all of the accelerated workloads, you need the GPU and that does massive parallelism deals with massive amounts of data. But to get that data into the GPU and also into the CPU, you need really an intelligent data processing unit because the scale and scope of GPUs and CPUs today, these are not single core entity. These are hundreds or even thousands of cores in a big system. And you need to steer the traffic exactly to the right place. You need to do it securely. You need to do it virtualized. You need to do it with containers. And to do all of that, you need a programmable data processing unit. So we have something called our Bluefield which combines our latest, greatest 100 gig and 200 gig network connectivity with ARM processors and a whole bunch of accelerators for security, for virtualization, for storage. And all of those things then feed these giant parallel engines, which are the GPU. And of course the CPU, which is really the workload at the application layer for non-accelerated apps. Great, so Perush, Kevin talked about needing similar types of services, wherever the data is. I was wondering if you could really help expand for us a little bit the implications of AI at the edge. Sure, yeah. So AI is basically not just one workload. AI is many different types of models. And AI also means training as well as inferencing, which are very different workloads. For AI training, for example, we are seeing the models growing exponentially. Think of an AI model like a brain of a computer or like a brain solving a particular use case for simple models like computer vision. We have models that are smaller, bugs have computer vision, but advanced models like natural language processing, they require larger brains or larger models. So on one hand, we are seeing the size of the AI models increasing tremendously. And in order to train these models, you need to look at computing at the scale of data center, many processors, many different servers working together to train a single model. And on the other hand, because of these AI models, they are so accurate today from understanding languages to speaking languages, to providing the right recommendations, whether it's for products or for content that you may want to consume or advertisements and so on. These models are so effective and efficient that they are being powered by AI today. These applications are being powered by AI. And each application requires a small amount of acceleration. So you need the ability to scale out or and support many different applications. So with our newly launched Ampere architecture just a couple of weeks ago that Jensen announced in the virtual keynote, for the first time, we are now able to provide both scale up and scale out, both training data analytics, as well as inference on the single architecture. And that's very exciting. Yeah, so if you look at that, the other thing that's interesting is you're talking about at the edge and scale out versus scale up, the networking is critical for both of those. And there's a lot of different workloads. And as Prash was describing, you've got different workloads that require different amounts of GPU or storage or networking. And so part of that vision of this data center as the computer is that the DPU lets you scale independently everything. So you can compose, you disaggregate into GPUs and storage and CPUs and then you compose exactly the computer that you need on the fly containerize to solve the problem that you're solving right now. So these new way of programming is programming the entire data center at once. And you'll go grab all of it and it'll run for a few hundred milliseconds even. And then it'll come back down and recompose itself constant. And to do that, you need this very highly efficient networking infrastructure. And the good news is we're here at HPE Discover, we've got a great partner with HPE, they have our M series switches that uses the Melanox 100 gig and now even 200 and 400 gig ethernet switches have all of our adapters. And they have great platforms, the Apollo platform, for example, is great for HPC. And they have other great platforms that we're looking at with the new telco that we're doing for 5G and accelerating that. Yeah, and on the edge computing side, there's the edge line set of products which are very interesting. The other sort of aspect that I wanted to touch upon is the whole software stack that's needed for the edge. So edge is different in the sense that it's not centrally managed, the edge computing devices are distributed, remote locations. And so managing the workflow of running and updating software on it is important and it needs to be done in a very secure manner. The second thing, that's very different again for the edges these devices are gonna require connectivity as Kevin was pointing out the importance of networking. So we also announced a couple of weeks ago at our GTC, our EGX product that combines the Melanox NIC and our GPUs into a single processor. Melanox NIC provides fast connectivity, security as well as encryption and decryption capabilities. GPUs provide acceleration to run the advanced AI models that are required for applications at the edge. Okay, and if I understood that right, so you've got these throughout the HPE product line, HPE's got long history of making flexible configurations. I remember when they first came out with a blade server, it was different form factors, different connectivity options. They pushed heavily into composable infrastructure. So it sounds like this is just kind of extending what HPE's been doing for a couple of decades. Yeah, I think HPE is a great partner there and these new platforms, the EGX for example, that was just announced, a great workload there is the 5G telco. So we'll be working with our friends at HPE to take that to market as well. And really there's a lot of different workloads and they've got a great portfolio of products across the spectrum from regular servers and one U2U and then all the way up to their big Apollo platform. Well, I'm glad you brought up telco. Curious, are there any specific applications or workloads that are the low hanging fruit or the kind of the first targets that you'd use for AI acceleration? Yeah, so the 5G workload is just awesome. We introduced with the EGX a new platform called Aerial which is a programming framework and there were lots of partners there that were part of that, including folks like Ericsson. And the idea there is that you have a software defined hardware accelerated radio area network. So a cloud ran and it really has all of the right attributes of the cloud. And what's nice there is now you can change on the fly the algorithms that you're using for the baseband codex without having to go climb a radio tower and change the actual physical infrastructure. So that's a critical part. Our role in that on the networking side we introduced a technology that's part of EGX within our ConnectX DX adapter called 5T for 5G. And one of the things that happens is you need this time triggered transport or telco technology, that's the 5Ts for 5G. And the reason is because you're doing distributed baseband unit, distributed radio processing. And the timing between each of those server nodes needs to be super precise, 20 nanosek. It's something that simply can't be done in software. And so we did that in hardware. So instead of having an expensive FPGA to try to synchronize all of these boxes together we put it into our NIC and now we can put that into industry standard servers, HP has some fantastic servers. And then with the EGX platform with that we can build really scale out software defined cloud ran. Awesome. Prash, anything else on the application side? You'd like to add in addition to what Kevin was talking about. Yeah, so from application perspective every industry has applications that touch on edge. If you take a look at retail for example there is all the way from supply chain to inventory management to keeping the right stock units in the shelves making sure there is no slippage or shrinkage. So to telecom, to healthcare where you're looking at constantly monitoring patients and taking actions for the best outcomes to manufacturing where you're looking to automate production detecting failures much early on in the production cycle and so on every industry has different applications but they all use AI they can all leverage the computing capabilities and high speed networking at the edge to transform their business processes. All right, well it's interesting almost every time we've talked about AI networking has come up. So Kevin I think that probably tees up a little bit why NVIDIA spent around $7 billion for the acquisition of Melanox and not only was the Melanox acquisition cumulus network very known in the network space for software defined operating system for networking but give us strategically does this change the direction of NVIDIA? How should we be thinking about NVIDIA in the overall network space? Yeah, I think the way to think about it is going back to that data center as the computer. And if you're thinking about the data center as the computer then the networking becomes the backplane if you will of that data center computer and having a high performance network is really critical and Melanox has been a leader in that for 20 years now with our InfiniVan and our Ethernet product. But beyond that you need a programmatic interface because one of the things that's really important in the cloud is that everything is software defined and it's containerized now and there's no better company in the world than cumulus really the pioneer in building cumulus Linux taking the Linux operating system and running that on multiple hardwares. So not just hardware from Melanox but hardware from other people as well. And so that whole notion of an open networking platform were committed to continue to support that and now you have a programmatic interface that you can drop containers on top of, cumulus has been the leader in the Linux FRR it's free range routing which is the core routing algorithm. And that really is at the heart of other open source network operating systems like Sonic and Dent. And so we see a lot of synergy here all the analytics that cumulus is bringing to bear with NetCube so it's really great that they're going to be part here of the NVIDIA team. Excellent. Well, thank you both much. I want to give you the final word. What should HPE customers in their ecosystem know about the NVIDIA and HPE partnership? Yeah, so I'll start. I think HPE has been a long time partner and a customer of ours. If you have accelerated workloads you need to connect those together. The HPE server portfolio is an ideal place that we can combine some of the work we're doing with our new amperes and the existing GPUs. And then also to connect those together with the M series which is their Ethernet switches that are based on our spectrum switch platforms. And then all of the HPC related activities on InfiniBand they're a great partner there. And so all of that pulling it together and now at the edge as edge becomes more and more important security becomes more and more important. And you have to go to this zero trust model. If you plug in a camera that somebody has at the edge even if it's in a car you can't trust it. So everything has to become validated, authenticated all the data needs to be encrypted. And so they're going to be a great partner because they've been a leader in building the most secure platforms in the world. And on the data center server portfolio side we really work very closely with HP on various different lines of products and really fantastic servers from the Apollo line of scale up servers to synergy and brilliant line as well as edge line for the edge and on the supercomputing side with the pre side of things. So we really work through the full spectrum of solutions with HP. We also work on the software side where a lot of these servers are also certified to run our full stack under a program that we call NGC ready. So customers get phenomenal value right off the bat and they're guaranteed to have accelerated workloads to work well when they choose these servers. Awesome, well thank you both. They're giving us the updates a lot happening obviously in the AI space and appreciate all the updates. Thanks Stu, great to talk to you. Stay well. Thanks Stu, take care. All right, stay with us for lots more from HPE Discover Virtual Experience 2020. I'm Stu Miniman and thank you for watching theCUBE.