I got to meet with Kevin Deierling to talk with him about what NVIDIA is doing and where they are going!!
NVIDIA Morpheus Video: https://ntck.co/3bxd9fh
Making of the GTC Keynote: https://ntck.co/3pYzaMl
🔥🔥Join the NetworkChuck membership: https://ntck.co/Premium
And video a company that is known for gaming graphics cards, right? Like I’ve got one of my computer right now from them, uh, 30, 90. Yeah. Be jealous. But I bet you didn’t know this they’re involved in something that I love even more than gaming, networking, cyber security. Yeah. And they’re not just involved. They’re kind of changing the game completely. And they’re creating new tech. That’s kind of crazy things like DPS or data processing units, which will help us do more with networking and cyber security and even AI than we ever thought possible. Sounds kind of cool. Right. Like hard to believe. So that’s why I invited the SVP of networking from Nvidia Kevin dearly to come on and tell us all about, because yeah, I’ve got some questions. Like what junk is this? How is it working? So buckle your seat belts, grab a coffee. Let’s dive into pretty much the future of networking. Kevin, who are you?
So I’m the SVP of networking at Nvidia. I joined through an acquisition of a company a little more than a year ago called Mella.
So I’ll be the first to admit. And a lot of my audience probably thinks the same thing. We didn’t know Nvidia was in networking. I mean, I’ve got my Nvidia graphics card behind me and my, my PC, but like networking. When did that come into play? So, um, that’s pretty cool. Uh, so why Nvidia getting into networking?
Yeah. So we’re now the largest provider of networking at the 25 gig and above speeds. We’re the number one internet provider at these high-performance speeds. And we also produce a technology called incentives and that’s the world’s leading HPC. So high performance computing and AI networking technology, and why it’s because you can’t build, uh, the, solve the problems and build a data center without networking and all of the problems can’t fit in a computer anymore. We have to think at data center scale, the data center is the new unit or computing.
There’s such an evolution of how that’s happened, right? So like I know with, um, I’m very traditional networking. I love my switches and my routers. I love my hardware. That’s vastly changed over the past 10 years or so to where it’s all software defined networking, we’re actually putting it on the server. Um, so a lot of people probably understand already, but I just had to say that. So we catch them up. We’re no longer dealing with network hardware. We’re at software on white boxes or just on standard or server hardware now, um, real quick, before we talk about what it is that you guys are doing, which is crazy, it’s the Bluefield do-good DPS and all that kind of crazy stuff. Before we talk about that, what’s the problem we have right now with SDN. What w what are we, what’s the wall we’re hitting, that’s preventing us from accelerating and doing.
Yeah. So, you know, this is kind of like goes back to Andrew’s since now, it’s almost a decade ago that he talked about software is eating the world and what’s hidden in that statement. It’s true. It’s totally true. It’s transforming industries, including networking and so afford to find networking as an example. So SDN is certainly an example of, but ultimately software needs to run on something. The point here was software to find networking is when they said, well, let’s just run it on what we have, which is a traditional CPU. And it turns out that you get all of these great benefits of software defined networking, which is flexibility, scalability. You can use, you know, modern software or continuous integration and continuous development techniques, the ICD, but you’re still running on the wrong horse, different horses for different courses. And in this case, the CPU is actually not very good at running software defined network,
Which is interesting. Cause I, you know, I come from a Cisco background and I know that when Cisco designs their switches, um, they have like a six chips designed just for switching packets, switching frames. And that always was in the back of my mind, like, okay, if we’re moving all this stuff to standard server hardware, are we missing something here? And I feel like that’s probably it, right? Yes,
Exactly. Right. So, you know, if you look, the traditional approach to this has been to develop a customer, but even for, for very large companies trying to develop a bespoke custom ASIC for a specific function, whether that’s a fire wall or a storage box or a software defined networking box is really hard. And you’re subscale. You’re not a semiconductor company. You’re not cranking out chips at the latest process node. Every single quarter, we push a ton of chips through this pipeline, this massive engineering investment that we make. And so we can build a chip that really is even better than a customer. So it’s software defined and hardware accelerated with our DPS.
Okay. Now that all sounds amazing. It sounds geeky. And I want to get really, really into the weeds on that. So explain to me what you guys are bringing to the table. Um, so we, we have the data center and we have our servers that are now running the network as well. It’s all just one big blob of everything on the same hardware. What are you guys adding to the picture to solve our problem of, like, we just can’t switch the packets fast enough. We can’t inspect things fast enough. What are you guys doing?
Yeah. So, you know, you described Nvidia perfectly. You said, Hey, you’re a GPU company. I know you guys is the graphics processing company. And over the last decade, what we’ve seen is the rise of the GPU to be able to do accelerated computing workloads, AI,
AI, and machine learning. That’s the big breakthrough. You know, people think about, uh, other stuff, crypto currencies, but it’s really AI and machine learning that has been the transformative trend AI as the most powerful technology force of our time. And so the CPU simply isn’t good at running AI and accelerated workloads. So instead of having the CPU centric server box, we really think about the data center is the new of computing with the CPU and the GPU to run AI and machine learning workloads. And now the DPU or the data processing unit to run all those things you just asked about data intensive packet, switching, packet, processing, encryption compression, all of those things. It turns out the CPU is really bad at running those. And instead of using a custom ASIC to do each of those functions, we can run software defined, networking, software, defined storage, and software defined security on the DPU. So really it’s all three together, the CPU, the DPU and the GPU that is building this new data center as the new unit of computing. So
I want to break cause like I know for a lot of my audience, that’s going to be like, okay, that sounds cool in theory, but like tangibly, what does that look like when you’re saying, okay, we’ve got, we know what the CPU does and we know what the GPU does, but then we have this DPU. Is that just gonna be like on the same chip set, the same motherboard as this server? Where’s that going? Where is that living at?
Yeah. So right now we’re just plugging that in to a server and we’re putting a computer in front of the computer. And so when you have a bunch of virtual machines, you actually need to switch between them. You can do that in software, running on the CPU and you get really lousy performance. And when we talk about performance today, you know, 25 gig is sort of the entry point a hundred gig, 200 gigs, soon 400 gig into every server are our big DGX boxes with lots of GPU’s in them. We have 1.8 terabytes per second of bandwidth going into that box, trying to imagine a CPU that’s able to process even just to look at the data, let’s say it’s a denial of service attack. If it has to look at every packet just to throw it away, it turns out you brought the CPU to its knees, running a DDoS mitigation, brings it to its knees. It actually accomplishes what you’re trying to prevent. And we can do that with zero CPU.
No, actually a another YouTuber, a friend of mine, David Bumble, he was testing 40 gig between two computers and he’s doing a file transfer test. And you probably guessed what the bottleneck was. It wasn’t the network. It was the CPU that just could not handle the file transfer. So yeah, data centers no different, right? Like it’s, that’s going to be the bottleneck. Um, so what’s, what’s the magic of the CPU. What what’s it actually doing?
Yeah. So it does software defined networking, software defined security and software defined storage. And it does all of those with hardware acceleration. So we have a bunch of CPU’s embedded CPU’s inside of the DPU. So we have a ton of arm processors and they’re super interesting. However, it isn’t actually the magic of the DPU. The magic is in those accelerators. So you talked about trying to run the file transfer. We celebrate storage. We do that with zero CPU. Look, we don’t even use the arm processors to transfer files. We use a technology called RDMA and we can move at line rate. We can move at a hundred gigabits per second or 200 gigabits per second files to file without involving the CPU at all.
That’s crazy. So my, okay. I’m trying to wrap my mind around this. So like we’re not involving a switch at this point. It’s just the DPU doing the job, offloading that and an SDN architecture. We’re talking like VM to VM transferring files and it’s at line rate. That’s what we’re talking about.
Absolutely. Absolutely. And so the DPU, that’s the great thing. We talked to ordinary switches, so we can talk to our own switches or other people’s switches and communicate across the entire data center at super low latencies. But the real magic we like to say offload accelerates and isolate. Those are the three things that the DPU does. Offload is simple. We just take tasks that we’re running on the x86 and we run them on the DPU. However, if we simply ran them on the embedded arm processors, we actually wouldn’t achieve the kind of Speedo that we talked about because the x86 cores are pretty good at doing what they do, but it’s that second piece. So we say offload and accelerate, we have accelerators hardware acceleration to do security and storage and networking. So when we move the files, it’s our accelerators that do that. And then the last piece isolation is really related to security. Okay.
Okay. Well, so let’s talk about security cause I saw some cool use cases, especially from one of the companies I’ve loved using in the past. When I actually had a real job, uh, Palo Alto, um, I saw they started to use this technology, this blue field. What’s the official name for this, by the way, I know it’s data center or DPS data center data center. What does it get in processing unit or?
Yeah, it’s a blue field data processing unit. Okay.
Okay. Gotcha. Um, so Palo Alto started doing this. W what, what are they doing with this now?
Yeah. So, you know, you asked what was the problem with SDN and all of the companies like Palo Alto, who’s a leader in building firewalls to look at all of the traffic. And it’s not just a traditional firewall where it was ports that you would block. It’s actually looking inside of the packets doing deep packet inspection. And when you start to do that, it turns out the problem with software defined networking or software defined security is that the CPU is not that good at it. And if you have a hundred gigabit per second link, you’ll bring the CPU to its knees at about 10 gig. That’s about what you can achieve. So what we did working with Palo Alto, we have this platform or software API called Doka and they were able to leverage those accelerators. So they took their software defined networking, and they did something that they call ITO, which is actually their offload.
So it offloads the traffic to the DPU and the DPU can inspect the traffic for them. So looking at all the packets and the headers and analyzing what the flow is and decide which of those things that the CPU should look at, which we can forward or redirect, or, or run a load balance or all of those things. And as a result, we get nine times better performance. So now we’re close to a hundred gigabits performance, and we used up zero CPU load to do all of that ITO, that acceleration that we’re doing for the firewall.
Wow. So like, I just can’t get my mind around that. It’s like zero CPU usage on that. It’s not hitting the CPU at all. It’s offloading all of that because I know that that’s an intense process inspection, especially at that level where you’re trying to look at all the different types of traffic encrypted non-encrypted, um, scan it for viruses and everything. Like that’s, that’s a heavy lift. Um, and you’re offloading that to that DPU. That’s a little crazy.
Yeah. So this is, what’s so great. You just nailed all the things that we do well, and it’s things that the CPU doesn’t do well. So if you’ll use a CPU to try to inspect the packets and look at where it’s coming from, or to decrypt the packet, we can bring a CPU to its knees running something like IP sec or TLS. And we do that with software defined hardware, accelerated with zero CPU load. And so I call about, I call it infinite efficiency. You get 10 times the performance or 50 times the performance with zero CPU and 50 times the performance divided by zero is an infinite efficiency game. Now I’m cheating a little because we were adding the DPU. But when we do the analysis, these parts pay for themselves, you use thousands of fewer servers in a mid-sized data center, and you’re getting the same efficiency from that data center.
So this is the thing that’s really great. We, we see the customers that are adopting this first are the hyperscale guys, the cloud guys, because this rich right down to their bottom line, what did they sell? They sell CPU cores to their customers. And if they’re chewing up 30 or 40% of their CPU cores running their data center infrastructure, their software defined infrastructure, that’s, those are cores. They can’t sell to their customers. So this rips right down to their bottom line, they’ve been the early adopters of this technology. And now we’re bringing it to the mainstream. It’s super exciting to be able to deliver this to all of the tier two cloud vendors, as well as enterprises. So really, really exciting for us.
It’s funny. It’s like, everyone’s like SDN. This is like the solution, the savior, like what we’ve been looking for. I didn’t anticipate this problem. And I was watching some of the videos on this and I heard a metric, like a third of the compute we have is dedicated to SDN. Like it’s sucking up all of that. I’m like, that’s, that’s a lot. So it totally makes sense to offload that to a DPU. So you have all that. I didn’t think about a cloud case like that, where yeah, they, they, they sell us the CPU course that that’s what they get off on doing. And, uh, goodness, that’s going to save a ton of money in that, that realm. That’s weird. Um, so you, you mentioned that though. Cause I that’s the question I had when I heard this deepest, um, Bluefield sounds like a very high tier, like people like AWS and Google cloud and Azure will be interested in this, but you said enterprise customers too. So what type of companies would have this in their own data center?
Yeah. So I think anybody who is looking at lots of data and has a requirement for service level agreements and what happens is that’s everybody today because AI, AI is every is everywhere. Every business is becoming data-driven and every business is becoming AI. So if you are running a workload like, you know, that requires speech recognition or you’re doing something at the edge and, uh, manufacturing, and you’re using robotics and cameras and analyzing images. If you’re doing anything here, we are communicating across a video link. Okay. People are streaming video across their networks in all of these different applications. It turns out you need to deliver packets really, really nicely. Precisely. We have something called the time synchronized data center called Kronos it’s part of our Doka package. And Kronos actually has precision timing built into the network cards. It turns out when you watch a movie at home and you’re streaming over the top, if you’ve unplugged, you need to get a file.
You’re just downloading a file. It’s called happens to be a movie. You need to download it not fast, but really slowly at precisely the rate that your home is able to receive and display the data and everybody’s home slightly different, and everybody has a different TV or whatever device they’re using to watch it. So now imagine trying to do that for a hundred thousand homes in software, where you’re counting out the packets and timing them. We have a technology that’s called packet pacing that will do that across a hundred thousand homes, very precisely, very accurately. So really any business you can imagine needs hardware acceleration. So software defined networking is fantastic, except it chews up all the CPU’s. You want to run your CPU’s to run business applications. That’s what your business is and all of that infrastructure. That’s what the DPU does for you. It accelerates it.
I love that example of the streaming cause like you, we all watch Netflix, we all binge, but you don’t think about how these companies can accomplish sending all those streams to all these thousands. Like I think Disney just announced they held like a million Disney plus subscribers. I can’t imagine the infrastructure behind that. I’m sure you can. Uh, but I can’t imagine what they have, uh, B uh, being able to timely deliver those packets, uh, is insane. Um, now what are some other things that are involved with, uh, you mentioned Doka like, so let’s talk about Doka a little bit, cause like I know we have Bluefield, uh, the DPS, but Doka is all about empowering developers to actually build on top of that. Um, and that kind of, I’m going to ask you two questions at once and you probably answered them perfectly. Um, so SDN, I know SDN to be like software that various vendors provide for us like Cisco Palo Alto. There’s a ton of them that do it now, are they going to be using things like Doka to be able to build on top of these DPS?
That’s exactly right. That’s how we’ll enable all of our partners. And we have a, an analogy in the GPU side of our business, which is called Kuda and Gouda is, is the framework that people can develop all of these AI and machine learning. And it really allows them to benefit from those GPU’s and the same Kuda programs that ran on GPU’s that are 10 years old, will run on our latest, greatest ampere today. So Doka is for the DPU. What code is for the GPU and Doka is this open SDK interface and a ton of libraries and reference applications that run on top of that. And what it lets us do is take things like our, uh, on the verse, which is a digital twin of whatever you want and run that on top of our GPU’s and our DPU is exchange data extremely quickly, very, very large data sets. The amount of physics that we run to create the Omni verse is just incredible. So Omni verdict is an amazing tool.
It sounds like you’re making all this up. It sounds very, Saifai like I’m reading a script from Marvel or something, the omniverse, um, speaking of Saifai, um, I was watching this video about your thing called Morpheus, which killer name I’m not that’s that’s cool. But I saw it was showing AI driven analytics and the cyber security. And so showing all these like strands of like fiber hair branching out and analyzing all these streams of packets and, and connections. Talk about that a bit. Like what does that doing?
Yeah. So Morpheus is a really cool cybersecurity application that benefits from the DPU and the GPU. And that’s what we’re really seeing is the synergy between the networking side and the AI side of the business. So if you think about what’s happening in a complex data center today, when people accidentally move from a development environment to a production environment, they can have their login credentials for, uh, you know, exposed to the entire world. And suddenly somebody can take over, you know, 10,000 virtual machines, somewhere in the cloud and start doing something that they should probably shouldn’t be doing that never happens. Never do. You never take your get hub and accidentally expose that, but you think about all the threats most of the time, it’s actually not being hacked. We have hardware that can detect malware and we can run that in the background constantly and detect that. But actually most of the time it’s just operational procedures that you inadvertently exposed somebody’s password or their credit card number or social security or health records or your own passports. And so it’s really hard in today’s environment to just look at log files. I looked at some of these log files. They’re just massive. And just, you can’t even imagine a human being figuring anything out.
Oh yeah. We have entire industries dedicated to log file ingestion. It’s crazy.
Yeah. So, so what we’ve done is automate all of that and we put AI on it. So instead of having a specific set of rules that say, I’m looking for this problem, you know, there’s always this question after there’s been a breach, it’s like, well, did we have a rule to look for that? We’ll know because otherwise we would’ve found it. AI, we can make these much broader rule sets. Hey, we’re looking for things that look like this. We’re not saying where they are or what, when they might occur. We’re looking for humans that are behaving like machines or machines that are behaving like humans, where things that are being sent that looked like they shouldn’t be sent. We’re not defining it precisely. That’s the beauty of a neural network and AI. And so the blue field is generating all this telemetry data on all of the packet information. And it’s sort of our camera into all of the network activity. You can think of it as a probe or sometimes you call it a span port and the networking moral. We actually are doing that in software, sending data to the AI. And then the AI is crunching in saying, Hey, this looks a little fishy. We’re going to quarantine this. And now you can start to actually open and close that aperture of what you want to look at. So really cool technology
That is intense. So I know like that that’s a big industry just being able to, uh, detect the issues. And then you’re, you’re so right. Like, um, the big hacks mostly are not like these genius hackers who just hack the mainframe, right? It’s mostly people making mistakes. And I mean, for anyone in infrastructure knows how easy it is to make a monumental mistake, just going from development to production. And that’s not just like programmers, that’s just, that could be a network engineer making a small change. Whoops. I love my credentials in plain text somewhere. Like that’s so easy to do. So my goodness has it like that, like in like, and I, and I saw one of the examples where it showed up, it was analyzing for encrypted traffic and it suddenly noticed that a normal encrypted traffic, uh, uh, I guess flow was actually un-encrypted randomly and like, okay, what’s going on there and found it, turned it red, alerted it. And then you can connect that to, um, other software like ticketing applications or remediation applications that can help the team solve the issue. That’s wicked. Awesome.
That’s great. That’s exactly right. Because, you know, we, we just call that good security hygiene and it’s really hard for human beings to just detect everything. And so what we want is you to sort of describe what that hygiene looks like, what the behavior is that you want, and then let a machine enforce it for you. You don’t want to be looking at Splunk log files. Swamp was somebody that we work with on Morpheus to you. Don’t have to look at just network traffic. You can look at log files and then detect all kinds of things. Half the time you’re looking at different log files, they’re breaking, they’re changing, trying to have a fixed set of rules to look at them is impossible. And so using AI to solve this problem is really just the smart thing to do. And it’s the only thing that’s going to work looking for. It takes an AI to fight cybercrime.
Yeah, absolutely. Now with, um, with Doka and Bluefield and all the DPS and stuff, we, we, we know what, like what we’re looking at now, what it can do, what do you think it’s going to do? What I think it’s going to do for the industry? Cause like I know your team is probably all about looking forward and seeing what can we do with this and developing for that. What do you think is going to happen?
Yeah. So you you’ve asked the perfect question for Doka because that’s really what it’s about. It’s about protecting our developers investment in applications. And so our goal with Doka here is to accelerate the innovation that our partners are doing on top of this accelerated framework for both the DPU and the GPU. And so, and now we need to protect that investment. So we’re going to have Bluefield too soon, Bluefield three Bluefield four, which adds GPU capabilities and AI into a converged platform. And what we want is that all of the programs that are being written today will run on those future programs. And today we just released the first version of Doka. We announced less than a year ago. We released the last one in April. We just released the one.one version of that as we go to one dot two and two dot, oh, we want all of those new capabilities and accelerations to be available for our partners. So you were asking about different companies that are going to develop storage companies, networking companies, security companies, ultimately people that are building the omnivores, this digital twin of anything, anything you can imagine, we will create a digital replica of it and allow you to simulate you wouldn’t build a things today without having that simulation environment, we’re simulating everything, including the data center.
I’m just trying to get my head around this. That sounds so crazy. Um, you know, as, as I learn about this stuff, I’m thinking like, what’s it going to change for like jobs, people looking into like becoming network engineers and all network engineers are kind of developing a developer skills, learning Python and learning development. Um, people who want to start developing on Doka the SDK, is it mainly focused on networking? Like, is this becoming like a new kind of job sector, like network developers who work on this kind of stuff?
Yeah. I think, you know, the great thing is, is when I started off, I was a chip designer and I was dealing with transistors. And at some point you jumped to a higher level of programming, uh, of chips. We designed chips, not at the transistor level, but we wrote a language and I see the same thing happening here with Doka and the Bluefield DPU. So all of that expertise that the networking and it administrators have gained over their careers is absolutely vital to understand what you’re trying to do, but it, it makes you more efficient. It lets you program at that higher level. And so even things like regular expressions, we can start to do regular expression searches. So if you’re using the Python regular expression, we can accelerate all of that with the Bluefield that has an engine that does that. If you’re looking for giant, uh, malware signatures for, for attacks, you can just program that library and we’ll give you a much higher level framework that you’ll develop on top of, but you need to understand what it’s doing to use that. And so we have a lot of activities now on training, you can go to our Doka and download it. And that’s part of what we do well is really engaged with our partners because our goal here is to help people to innovate faster.
That sounds really interesting. Um, now I saw this, I was watching a bunch of the videos on the stoke and stuff and, uh, and DPU is, and I saw that you, you have this Nvidia platform where it’s like cloud gaming and I saw that you’re using, or you’re about to start using deep use with this. Can you, can you talk a bit more about that?
Yeah. So what we have is what’s called our G force now platform and what it is, it’s a cloud gaming platform. There’s millions of users that can log in. So instead of needing to own the best games for every single copy that you might want to run, we let you do that in the cloud. And so you get a service. So thinking of it as gaming as a service on the cloud. And so we have these killer GPU’s that we build into our cloud, and then we let our customers log in. And again, now you’re talking about an interactive environment where, you know, you’ve got a joystick and you’re playing a game and we read, need super precise, uh, quality of service. We need low latency and all of that, but we have to have anything that’s exposed to the internet. You have to be able to protect against people that are externally facing because you know, there’s going to be all sorts of attacks, anything that’s facing the internet sooner or later, it’s going to face and you need to be able to mitigate
Now with, um, I know cloud gaming. It’s, it’s been like a cool concept, but I know a lot of people are like, ah, it’s kind of a weird thing. Cause I know we’ve had things come out in the past, like 10 years and people have tried it like Google stadia or something. And it hasn’t been very well received. It’s like there’s latency and it’s crappy. Um, how will DPU use and all the things you’re doing with, uh, G-Force now helping with this?
Yeah. So one of the things that we think is interesting about G-Force now is it really is the same sort of experience we’re giving you a seat on the server. And so other people are doing things with, uh, you know, not really giving you access to the server. And in our case, we’re doing that for creating a VM and you get that via and you’re running on the server. And so there’s, there’s security issues we can mitigate with the DPU when we, when, when we do that, because we put a computer in front of the computer and this is this offload accelerate and isolate by isolating, we put all of those infrastructure services in a different domain than the application processing domain. In this case, the application is the game itself. And now we have some infrastructure underneath that. That’s independent of that. And so that isolation piece is really, really important to be able to protect our infrastructure. And at the same time, we’re really giving the gaming user the same experience that they’re used to because they’re loaded that game is loading on the machine that they’re used to.
You know, it’s so funny as you’re saying that it’s like, it it’s so funny cause it’s like, we’re so focused on hyper converging our network a little bit. Like we we’ve gone from having it displaced on different, different S uh, hardware to having a one piece of hardware. And now we’re looking at, okay, it’s too much, let’s separate it and software now let’s, let’s give them some space, give them some breathing room. That’s kind of the direction we’re going now. Right?
Yeah. That’s right. So, and it’s that you you’ve gotten hit on exactly. A super interesting point about the isolation. I had talked about the security benefit of having separate infrastructure and application processing domain. So you’re running your application on the CPU and the APU, and then your infrastructure is running on the VPU that gives you this security benefit, but it also gives you an operational efficiency benefit. And as an example there, now I decoupled my applications from the underlying infrastructure and it may be that, you know, in our case we designed chips. We have a lot of EDA software for synopsis and cadence and mentor to design chips, and they have dependencies on the operating system. And so we may want to add a new Verner version of the Lennox carnal, and we can’t do it because of the application is dependent on a specific version of the kernel.
When we decouple these, now we can actually run at different paces for our applications and our infrastructure. So if there’s a new security patch that we want to put on the infrastructure side, we do it, it’s independent of whatever, the applications that are running. We just eliminate all of those dependencies. When we talk to our own CIO, that’s what really excited him is the operational efficiency that we gained by decoupling. So when you mashed everything up with SDN, it turns out there’s some downside of that performance, but also you end up with those tie in between the application and the infrastructure. And now we’ve decoupled that again.
Ah, so it w when you, when you’re saying that I just screamed to me microservices, right? Like we everyone’s containerizing things and separating things, but you’re doing this at more of a, uh, more of a hardware level, but all in the same area. Um, that’s interesting. That’s really interesting. Um, I did see something about a partnership with VMware and Dell and VMware, uh, cloud foundations. Uh, can you talk, I, it’s probably a rehash of what we already talked about, cause I know it’s, it’s the same stuff. It’s your deep use on the server offloading isolating. Um, is there any different with the VMware Dell relationship?
Yeah, so VMware is a great partner and obviously Dell on the hardware side. And so we really did two different things with VMware. First of all, it was on the AI side. Historically our AI has been available, uh, running as accelerated computing on the GPU, but primarily for the sort of Linux environment, what we did was the announcement of what we call MBAE, which is our, uh, platform that runs on BM ware. So it’s all of our AI in a virtualized VM-ware environment. And that’s super important because really being aware is the operating system of the enterprise. And so now all of a sudden you can take your VMware environment and the way you virtualized things and manage things. And now you can run AI applications, all the Nvidia AI applications that we have, whether it’s Clara or healthcare or Isaac for robotics or natural language processing with our Rebbe platform, metropolis is our smart cities platform. All of that can now run on VM-ware. So that was the first thing we did. And the second was called project Monterrey, which is taking that offload, accelerate an isolate, taking the VMware environment and putting it on to our Bluefield CPU’s. And so together, the two things are super powerful. You get AI and you get this an offload accelerate and isolate in a familiar VM-ware environment.
Okay. Okay. So I have a few questions that, cause that sounds so cool. First of all, the AI stuff is like, it’s over my head, but it’s amazing because like, I know what it does, but I don’t know. Also, I don’t know what it does. Um, secondly, the VMware side, so I’m very familiar with VMware and how you can carve out and use CPU’s, um, and share them amongst your VM storage, networking, everything. How does that work with DPS? W what’s it look like, uh, when you’re actually like using, uh, ESI and vSphere and everything.
Yeah. So that’s a great question. And the answer, I think you’ll like, I hope so exactly. Like you’re used to, it looks like, it looks like it’s XiO it looks [inaudible] and it looks exactly like the environment that you’re used to using and all we’ve done now is really take the things. Whether you’re going to encrypt the channel, you’re going to make the same function call to say, Hey, I want this to be encrypted. I’m going to spin up a VM and I want it to have access to storage. And I want be set. I’m using some Tansu containers, all of the things that you’re familiar with in the BMI environment, we’ve just mapped those onto the DPU and where we have an accelerator, you get massively better performance. So it really gives you the best of both worlds. It gives you that familiarity of VMware plus the accelerations that we get from the DPU.
Dang. Okay. So the thing we’re talking about that, you know, cloud providers are interested in that we throw up in the cloud and like, they’re, they’re just doing this amazing DPU stuff, offloading everything, and we can do the same kind of stuff on like Dell hardware. We can put it in our own data centers and it’s got a DPU on it and we can use ESI, vSphere, whatever you want to call it. And it’s kind of transparent. Like when we incorporate networking V Sans or we’re going to be motioned something that’s, that’s using the DPU, we just don’t realize it.
That’s exactly right. And so one of the challenges there is to make all of that just work transparently while at the same time taking advantage of the hardware. So we have this technology called RDMA or Rocky RDMA over converged data, and we can transmit data without using the CPU at all. We do all of the things that normally TCP IP does in software with packet and sequence number checking, and CRCs and all that stuff. We just do it in hardware. But to make all of that work transparently, you know, to make the motion work over audio may to make the sand work over, you may, it’s a ton of work to actually make all of that plumbing take advantage of the hardware accelerators. But when we knew you get this magical experience, everything just works faster, better, more reliable. You don’t get these long tail latencies. What, what Google called the tailored scale? Uh, we’re doing it very deterministically in hardware. So you don’t have the situation that software most of the time, it delivers the data with one millisecond latencies, but every once in a while you see 50 milliseconds of latency because something’s happening
Right now. I’m just trying to think about how can I implement this into my home and my video editing infrastructure. Like that sounds amazing. I want to take advantage of that. Uh, that’s a, that’s a different price point than I’m used to though. Um, that’s amazing. Uh, so I think we’re, we’re getting pretty close to we’ve covered everything I PR I, I think as we, I think I’ve missed out Kevin, that we haven’t covered thus far.
Yeah. Well, you mentioned your own, uh, video environment. You know, media and entertainment is just loves this stuff because the world is moving from what serial digital interface, these sort of bespoke custom interfaces for digital to internet. So now we have something called SM PTE, which is driving the 2110 on top of the ethernet. And so as a single 4k or AK stream can just completely gobble up a 10 getting link. We’re talking 2,500 gig links. And it turns out that when you’re doing Hollywood production of video, for example, you need really precise timing. You may have tons of different cameras feeding in and they all need to be synchronized super tightly. And so we’re doing that with the technology. We have cold river max, as part of our Doka platform. We just see so much excitement here because everything is going to be networked. Everything’s going to be distributed microservices containers. And when you’re doing that, the data centers, the new unit of computing, you need to put it all together. It needs to be deterministic and precisely timed. That’s what people are used to in the computer. Now we need to do that in data center scale. And we think you need a DPU to do that.
Yeah. So many things you just said that I need to unpack. So, so I didn’t know that we’re doing V like production, like level of like red cinema cameras that are somehow connecting to either net and we’re transporting video.
Absolutely. It’s very cool. It’s really exciting.
Yeah. I love networking. All the things. That’s really neat. Like it’s like, I’m like, yeah, let’s run power over ethernet, like power our houses with Ethan that, and now it’s connect everything with Ethan. That that’s, that sounds amazing. So we’re saying, looks like a Hollywood production. They got multi-camera setups, obviously. Like you got all kinds of guys doing all kinds of stuff, but it’s not worked together with a couple I’m more than a couple of hundred gig links and we’re having to use obviously deep use in Doka and everything to, uh, manage all that to actually make it work without it just dying.
Yeah. And you know, as always, it’s the management as well. That’s the tricky part because if it takes forever to set it up and you can’t figure it out, then that doesn’t work. So just the discovery and all of the protocols, and that’s, what’s great about the DB. We can run that on the DPU and it says, Hey, I just plugged in a new monitor or a new camera. And then you have these agents that are actually saying, what’s going on. You discover everything, you figure out what it’s part of. You can isolate things and create clusters, and then you can bring it all back together. The, uh, studio of the future is really based on internet and software defined networking and the Bluefield DBU accelerates at all.
It’s so crazy. It’s like, you wouldn’t think that I’m like, I love that being an it, it can pretty much take you into any industry, but I didn’t think about it being like that close to the filming of a movie. Like you’re right, you’re right there. That’s where the networking’s happening. That’s amazing. Um, and then containers and stuff. So I, I’m going to plead some ignorance here because I don’t understand, uh, when you’re saying like neighbor discovery and stuff, I started thinking of like networking protocols, like OSP F ER, GRP. Are we talking about that when we’re talking about Doka and DPS, or are we talking about something else?
Um, you know, it’s not an autonomous thing. It’s more like software defined networking and the original vision of software defined networking is you had centralized management. And so you put something in and it goes, talks to its manager and it says, Hey, I’m here and I’m a monitor or I’m a camera. And so, and the great thing about when you look at production, if you’re a Hollywood studio, when you’re producing something, then you need, as you’re editing it, it’s all one compressed. It’s all raw. It’s all massive files. Every single frame has to be uncompressed. So the amount of data that you’re moving back and forth as you’re editing this people are they need the full uncompressed frames. And then later, when you finally edited that you run through that depression phase, and then you’re finally gonna stream it across. And that’s like, for example, one of the other platforms to somebody’s phone, but all of that, you know, when you’re dealing with uncompressed video, you need super high bandwidth and you need it to be automated. You can’t figure out, oh, you know, where’s my camera. What was that? Camera’s IP address? It needs to be automated. They plug in the camera. It tells somebody, Hey, I’m here. What do you want me to do?
Oh gosh. Yeah. I need to get my hands on some of that for my live streaming. Can I be the first YouTube or that does that sounds intense. Wow. Okay. Okay. I just, man, I didn’t think that’d be the application. That’s really cool. Um, so that was a neat one. Anything else that I’ve missed? Cause like this is like a big new world to me. I don’t want to miss anything. That would be really cool in the video.
Yeah. So I think the other thing is you should look at what we did with our keynote and we just published sort of a GTC keynote behind the scenes. And one of the things there is we have our Omni verse, which is this digital twin. And it turns out we created a digital twin of Johnson, the CEO of Nvidia, and we create a digital twin of his entire kitchen and then put it into the GTC keynote. And nobody noticed it was all AI generated.
Wait, hold on. You, you mean that? Hold on a second. I was watching that keynote. That guy was AI
That video in his kitchen. So go back, I, then we could share the link. It was, there was a part of that where we literally scan Jensen. We created a CG model and it was AI Jensen. Wasn’t the real Johnson. It wasn’t even as real kitchen. All of the stuff in there was being generated and everything has motions has, has face. Uh, gravity is being modeled in Omni verse. That’s what our neighbors is. It’s a digital twin to the point where frankly, I saw it.
Can we just call it the matrix? I mean, that’s what we’re talking about right now. Aren’t we let’s be honest here. Uh, that’s insane. I need to go back and watch that now. Cause yeah, I mean, I, I watched a lot of the keynote and then I would not, I didn’t skip a beat on that. It just all seemed normal. That’s hilarious. What a great prank, by the way, like that’s like you guys even notice AI, the omniverse you were there,
We were there and didn’t even know. And what’s really neat about it is that the BMW factory was another thing we showed in the real people interacting in a virtual world, you had real people that had XR. They had VR and they had sensors on them and then they were acting and they were acting in a virtual world. So they projected themselves with a headset into this virtual world and they were picking up things and figuring out things about the BMW factory floor that, you know, rather than build a billion dollar factory and then find out it’s all messed up and you need to redesign it or a building, do it in the virtual world. First figure out, optimize there and then build the real thing. And so this is the future. You asked them what the future was. The future is. There’s going to be a digital twin of everything and that’s what the Omni versus, and we’re super excited about it.
That’s so Saifai oh my gosh. That’s so weird. My gosh, a digital twin and Omni verse. I mean, I just, all the S it’s like all the scifi movies we ever watch is coming true. Um, now do you see like, like I know obviously this is going to have a huge, um, enterprise, uh, use case for so many different, uh, um, areas of a company’s everything, but do you see it like hitting the consumer base? Like I know AR VR is a huge thing. Do you see this like accelerating our descent into not living in the real world anymore?
Yeah. So I think it’s going to augment the real world and I think that’s the key of augmented reality. I really do believe that that, you know, look the, the real world’s a great place. Uh, but there’s some things that you just can’t do yourself for whatever reason. I mean, we saw things where, where people that are unable to, to speak, for example, they’re connecting up to their brain, suddenly they’re able to speak. There is an example of augmentation. What we’re really challenging people to do is think of things that human beings can’t do today that we’d like to be able to do. And let’s go do that together should be fun.
I there’s just so many things that the supplies too, it’s like, it’s, it’s, it’s mind boggling because when you talk about offloading and accelerating and what was the other word, it was, uh, offloading, accelerating, isolating. That’s what it was. It’s so many industries. And, um, a lot of things I’m excited about, like cybersecurity is a big one, VMware like that’s, that’s really cool, but then things I even think about, um, cinema, uh, the harmony versus what the Chuck that’s that’s so cool. Um, anything else, like any closing thoughts on this Kevin cause like, I mean, you’re, first of all, I’m impressed by the sheer breadth of knowledge in so many industries that and how this applies to everything. Cause it really touches everything from VM-ware to knowing about V San and vMotion to movie production. And anything else that we haven’t covered that you might want to close with.
Yeah. So I think that you nailed it. We think every business is going to become an AI business. And that’s not a statement that’s actually particularly bold to make, because all I’m saying is that the businesses that don’t become AI and embrace it, aren’t going to be around anymore because AI is the most powerful technology force of our time. And the companies that embrace that and start to deliver new services to their customers and be able to operate more efficiently and to use their data more efficiently. They’re going to thrive at the expense of other companies that don’t. So the reason that it’s so broad is because every business is going to become AI. It doesn’t matter whether it’s healthcare or retail or, you know, robotics, manufacturing, distributing self-driving cars, all of those things are going to become AI. So we’re super excited and I’m learning a ton all the time from the people at Nvidia that are focused on all those different areas. And really our goal here is to accelerate other people’s ability to innovate and all of our partners and our customers. So it’s an exciting future. We need to make a digital twin of a Utah,
Please. God, I worked too much as it is. My wife would love that actually. Yeah. A little digital twin of my whole production here. That’d be so cool.
Well, Kevin, I so appreciate you coming on and talking about this crazy stuff. I mean, it’s really [inaudible] we talked about creating the matrix. Um, we talked about solving cyber crime with AI and robots. We talked about filming movies with 100 gig links. Like that’s, that’s so crazy to me. I love that it can touch every area of life. And I mean, talking about how AI will be the standard for companies now, like if you don’t embrace it, then you’ll get left behind that’s that’s crazy. And I know people think like AI, like, okay, most people are like, it’s a negative thing. It’s it’s robots taking over jobs and taking over things. And eventually we got sky net. That’s not, that’s not the case. It’s just accelerating and doing things that, like you said, humans can’t, um, it’s just enabling us. Yeah.
It’s like any other sort of technology innovation it’s really going to liberate people. And so I think for the people that are actually looking at their jobs, they should be thinking about, Hey, how can I use this? Are there things about my job? I don’t like that I don’t want to do. And how can I leverage this technology and AI to start to do things that I do enjoy doing. So we think that it’s actually going to be creative. Uh it’s it’s an example of, you know, what, what we’ve seen before with technology. And, uh, we think that there’s a tremendous amount of business and creativity that’s going to occur because of AI.
Oh yeah. I, I totally agree. And I can’t wait to see what happens. It’s going to be kind of nuts anyways. Thanks Kevin for coming on.
Thanks. I enjoyed talking to Josh.