In The Arena by TechArena - Delivering a Foundation for AI, 5G and Edge with NVIDIA

Episode Date: February 16, 2023

TechArena host Allyson Klein chats with NVIDIA’s Rajesh Gadiyar about his company’s strategy to accelerate 5G and edge adoption including cloud native vRAN.

Welcome to the Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now let's step into the arena. Welcome to the Tech Arena. My name is Alison Klein, and today is going to be a great episode. We've got Rajesh Gadiyar, VP of Telco and Edge Architecture from NVIDIA with us. Welcome to the program, Rajesh. Yeah. Hi, Alison. Good morning. It's great to be here today talking to you. It's fantastic to hear your voice. Why don't you just start and introduce yourself and

your role at NVIDIA and how it relates to the topic of the day, which is 5G deployment. Yeah, awesome. So at NVIDIA, I'm vice president of telco and edge architecture, and I'm relatively new at NVIDIA. I took on this role in September. A key focus area for me here in this role is 5G and the radio access network and the RAN virtualization, including cloud RAN design and development. I'm also looking to help communication service providers or the telco operators modernize their infrastructure and accelerate the transition to a modern software-defined cloud based on accelerated compute infrastructure so they can reduce their costs and improve the utilization of their infrastructure. Now, as you know, before NVIDIA,

I was at Intel for many years. I was the VP and CTO of Intel's networking business. My team built DPDK, the Data Plane Development Kit, that enabled high-speed networking on standard server platforms and the resultant disaggregation of hardware from software. And much of that work also enabled the NFV and SDN transformation for the last decade. Now, more recently, I led the 5G platform architecture, including the development of key cloud-native technologies for network applications. Now, I'm new at NVIDIA, but I'm super excited to be at a company that is changing the computing landscape. We are innovating and delivering accelerated compute platforms for AI and cloud computing and also networking in 5G. And by the way, it's great to connect with you again after a couple of years.

So it's good to be talking. Now, Rajesh, the topic, as I said, was 5G deployments and the status of 5G. Obviously, 5G has been talked about in the industry for a long time, and you have been in your various roles very involved in the development of 5G. Where are we at today, in your mind, with the deployment of 5G around the world? And how is the confluence of 5G, cloud computing, and AI coming together in your mind? Yeah, that's a great question, Alison. The 5G deployments have been accelerating around the globe, and many telco operators have already

rolled out 5G services, and they're expanding rapidly. In addition to the telco operators, there is significant interest among the enterprises to use 5G to set up their private networks, leveraging the capabilities of 5G, like higher bandwidth, lower latency, new technologies such as network slicing, and then millimeter wave and CBR spectrum. Now, in my view, the 5G build-out comes at an interesting time. If you look at over the last two decades, cloud computing has matured and has become the playground of choice for developers to build their applications. Cloud offers many advantages, mature software tools, automation orchestration, business agility, and more TCO.

Furthermore, applications in every segment, such as industrial robotics, cloud gaming, smart cities, autonomous driving, smart farming, they're all increasingly using artificial intelligence, AI, to enable transformative experiences. This confluence of 5G, cloud computing, and AI is super exciting. In my view, it will drive many new innovations over the next decade. But I would contend that 5G radio access network is somewhat of a weak link today, and it has not kept pace with the AI and cloud computing. We need to innovate faster and build a more performant, scalable, programmable, and automated radio access network. And I think the timing is right. Virtualization of RAN,

the technology is maturing, and now we are at a tipping point to drive faster adoption of VRAN, virtualized RAN, and perhaps even move faster towards deploying the entire radio access network in the cloud, which I'll talk about later. Now, a quick word about NVIDIA. We have a strategic investment in 5G RAN platforms. We've developed a platform called Arial. And the NVIDIA Arial platform is a software-defined full 5G layer 1 offload that's implemented as an inline acceleration in an NVIDIA GPU. So this NVIDIA Aerial platform, which we'll talk about some more in this podcast, is a key technology foundation for building virtualized RAM. And it implements all the 3GP and O-RAN compliant interfaces. So at NVIDIA, our goal, therefore, is to deliver a full platform with cloud-native software that serves as a foundation for 5G, AI, and edge applications.

Now, I'm glad you brought up VRAM because it's been a big area of focus in the industry and somewhat of a holy grail in terms of being able to actually virtualize that radio access network. Can you explain the motivation behind wanting to virtualize RAN and what is the status of disaggregation of RAN? This is a great topic. There is a lot of excitement around 5G and the economics of deploying 5G has been challenging, particularly on the radio access network side of things. So 5G is driving a significantly higher RAN capex growth as compared to the previous generations of wireless LTE and 3G. The number of cell sites for 5G are expected to nearly double over the next five years. And consequently, the RAN capex as a share of overall TCO is increasing from 45 to 50% that it used to be up to 65%.

So it is also well known that traditionally the RAN is provisioned for peak capacity, which leads to significant underutilization of precious compute resources. The bursty and time-dependent traffic means many traditional RAN sites are running at below 25% capacity utilization on average. So as a result, it is really important to disaggregate the radio access network and drive more centralization and drive better utilization by pooling the RAM resources, the compute resources in the radio access network. This is where the work that ORAN Alliance is doing,

the Open RAM Alliance initiative to disaggregate traditional radio base stations into what are called as the NRUs, the remote radio units, the virtual DU and virtual CU instances with very defined interfaces between them. That's great progress. And it's resulting in a larger RAN ecosystem with more vendor choices. So as our industry accelerates the 5G deployments, scalable and flexible solutions are very much needed to realize the full business value of 5G. So disaggregating RAN software from the hardware and making the software open and automated for deployment on private, public, or hybrid cloud infrastructure. redundancy, and it's optimally designed for mobile network evolution over the next few years, including the next generation radio technologies, such as saving the path to 6G. So to answer your question, RAN disaggregation, centralization, and cloudification are inevitable,

and we are seeing some good progress in that direction, but it could be faster, and it needs to be faster. And by the way, like I said, NVIDIA's Arial platform is fully 3GPP and O-RAN compliant and it's a great solution for the virtualized RAN and Cloud RAN deployments. The solution is mature and we are actually driving some field trials at the moment and commercial deployments later this year. Now you've mentioned Arial a couple of times.

Can you tell me a little bit about how that was designed and is this a cloud-native architecture from NVIDIA? Yes. So, Arial is a full platform for 5G virtualized RAN and cloud RAN deployments. It utilizes NVIDIA converged accelerators with our Bluefinet DPU, the data processing unit, and A100 class GPU. Provides full RAN layer one inline acceleration and offload. It's also software defined and supports all configurations from 4D4R to massive MIMO 32T, 32R, and 64T, 64R configurations. So unlike other solutions in the market that hit one sweet spot, like 44R, but it doesn't necessarily scale for 32T, 32R, and 64 to 64R, the NVIDIA Aerial solution is sort of like completely software configurable.

So the same platform can be configured in many different ways, depending on the use cases. In addition to what I said about the Aerial platform, the Aerial software stack is designed ground-set as a cloud-native software. So the Aerial architecture facilitatesitates RAN functions to be realized as microservices in containers orchestrated and managed by Kubernetes. Now this modular software supports much better granularity and increased speed of software upgrades, releases, and patches, independent lifecycle management following DevOps principles and CICD, independent scaling of different RAN microservices elements, and application-level reliability, observability, and service assurance.

So for a true cloud-native RAN experience, the cloud, the edge platform, and networking, they all need to evolve. And in my view, there are some requirements that are critically important for the cloud-native and containerized RAN software stack to be commercially deployable. So things like time synchronization, CPU affinity and isolation, topology management. You always need a high-performance data plane and lower latency, quality of service guarantees, and high throughput, zero-touch provisioning, and so on. So one other thing is, as you know, the Kubernetes framework allows for something

called as operators that allow you to discover the acceleration capabilities and schedule workloads at the right nodes. Because if you do that, then that gives you a better performance for what and performance per dollar. So NVIDIA in our aerial platform, we've developed two key Kubernetes operators, the NVIDIA GPU operator and the NVIDIA network operator for VRAM deployments. So as you can see, the NVIDIA aerial platform, it's built ground up with microservices, cloud-native architecture, and it provides a solid foundation for building and deploying the 5G RAM completely in the cloud. That's really cool. You mentioned edge, so I've got to bring it back to edge. And

we've been debating on Tech Arena, the definition of edge. And it seems funny because we've been defining edge, I feel like, for years now, but I don't feel like we've coalesced around a definition. So I'm very interested in what NVIDIA's vision for edge is and how you look at defining edge. What are the challenges of growing that edge footprint and how NVIDIA plans on investing and engaging in the edge? Yeah, so let's unpack all of those questions. Let's first start with the definition of edge, right? So I think the need for edge computing

is fairly well-established at this point. We need the simplicity, composability, and automation of the cloud's native architecture, but we also need to support distributed processing. What I mean by that is processing closer to where the application is and where the data resides. So moving everything into a public cloud, that will just not work for today's latency-sensitive applications that need faster decision-making based on AI and machine learning algorithms. So the cloud fundamentally has to be distributed, and it has to come closer to where the application is. And to me, that is the essence of what

edge computing is all about. So it's not so much about location, it's really the flexibility and the scalability of applications, right? Distributed edge applications. Now, if you look at the computing and connectivity landscape, AI is becoming very pervasive and we are seeing a tremendous growth in AI and machine learning in every application segment, including many edge use cases and applications. But if you look at the compute performance, it hasn't really kept pace and Moore's law has reached its limit. So what we really need is an accelerated computing infrastructure that can keep pace with the needs of modern applications. Now, similarly, if you look at wireless connectivity, right, going from 4G to 5G and eventually to 6G, we need 100x or more generational improvement in increasing performance

and reducing latency. It's difficult to deliver this in a standard CPU-based implementation. So what's happening? So as a result, what we are seeing is some vendors are building fixed function acceleration in ASICs to supplement the lack of CPU performance. In my view, that's a completely wrong approach, and it sets us behind many years to the old era of fixed function appliances. This is the very problem that we've been trying to address with RAN virtualization on standard COTS platforms.

If you think about it, the whole virtualized RAN is that hardware-software disaggregation. How do we build radio access networks on standard server platforms. Now, this is where NVIDIA's GPU come into picture because an accelerated general purpose compute platform with NVIDIA GPUs delivers where more slow cannot, and it can be a great solution for both AI and 5G applications. So that's the first changing compute landscape and why we need the accelerated computing infrastructure. Now, there are a few other challenges that need to be addressed for the edge computing infrastructure. Now, there are a few other challenges that need to be addressed for the edge computing to become pervasive.

The biggest, in my opinion, is how do we provide an easy button environment for developing edge applications? And this is another area that we at NVIDIA have been working on. So NVIDIA AI Enterprise with our base command and fleet command software enables the enterprises to run their AI applications

in the NVIDIA GPU cloud, leveraging all the pre-built and hardened software for various vertical segments. I'll give you some examples. NVIDIA Metropolis, for example, for video analytics and IoT applications. NVIDIA Merlin for recommender systems. NVIDIA iSEC for robotics.

And NVIDIA Nemo for natural language processing, speech recognition, and text-to-o for natural language processing, speech recognition, and text-to-speech synthesis models. Think about how powerful it will be for the 5G connectivity to be available as a containerized solution for enterprises to deploy in the cloud on the same infrastructure that runs all these AI applications. That will be truly game-changing. That will transform how the world thinks about wireless connectivity. It will truly make 5G a cloud-based service that can be deployed on demand. That is our vision. That value when it's run fully in the cloud. In this context, though, we see a lot of challenges with RAN due to timing, synchronization, latency requirements. What's your perspective on how we solve that and how does Cloud RAN fit in?

I'm really passionate about this topic of Cloud RAN. After I joined NVIDIA in September last year, I've been leading the Cloud RAN architecture at NVIDIA. I have to say, recently I've been observing that there's been a lot of discussion in the industry about Cloud RAN, and as the industry leader in accelerated computing platforms and cloud computing, NVIDIA has been at the forefront of cloud RAN innovations. However, my observation is that many industry leaders are using the term cloud RAN to simply mean a cloud-native implementation of RAN. Now, while the use of cloud-native technologies or building RAN solutions is stable stakes and it's much needed, the real question is, does cloud RAN

just equate to using cloud-native technologies? And I contend that it is not. The real question is, does cloud RAN just equate to using cloud-native technologies? And I contend that it is not. I truly believe that a cloud RAN has to have all compute elements, the virtual DU, the virtual CU, and the distributed UPF, all completely deployed in the cloud. So therefore, from an NVIDIA perspective, we are changing the nomenclature, and we are encouraging the use of the term RAN in the cloud instead of Cloud RAN to describe 5G radio access network that is fully hosted as a service in a multi-tenant cloud infrastructure. Now, you may ask, why is this distinction important, and what is the motivation for RAN in the cloud? So, like we discussed earlier, RAN constitutes the biggest CapEx and OpEx spending for telecom operators. And it's also the most underutilized resource with most radio base stations typically operating below 50% utilization.

Moving RAN compute completely into the cloud brings all the benefits of cloud computing, pooling and higher utilization in a shared cloud infrastructure, resulting in the biggest CapEx and OpEx reduction for telco operators. Now, Cod's platforms with GPUs can also accelerate not just 5G RAN, but it can accelerate Edge AI applications. And telco operators and enterprises today are already using NVIDIA GPU servers for accelerating their AI applications. Also, it gives them an easy path to utilize the same GPUs for accelerating the 5G RAN connectivity in addition to their AI applications, which basically means it reduces the TCO and provides the best path for setting up enterprise 5G networks. In addition to all this, cloud software tools and technologies have also matured over the years and are now delivering the benefits of at-scale automation, reduced energy consumption, elastic

computing, and auto-scaling on demand, in addition to better reliability, observability, and service assurance. So overall, the value proposition really is, how can we shift CapEx to OpEx and make the RAN connectivity completely as a service offering in the cloud? The OpEx benefit that are delivered with auto-scaling and energy management kind of capabilities, and the overall TCO benefit because of multi-tenancy and using the GPU-based accelerated infrastructure,

not just for RAN, but also for running the AI applications. One last thing, actually, like I said earlier, some VRAN vendors in the market today are designing ASIC-based fixed-function accelerator cards for RAN layer 1 offloads. Now, ran built on these ASIC based accelerator is akin to a fixed function appliance in my mind. It can only do ran processing. It is a wasted resource when it is not being used like in the nighttime weekends when the utilization is low. So the NVIDIA's aerial platform with general purpose GPU accelerated servers deliver a truly multi-services and multi-tenant platform, which can be used for 5G RAM, enterprise AI, video services, and other applications deployed in the cloud with all the benefits that we talked about.

That's really interesting. And I'm so excited that we are entering Mobile World Congress season and we're going to see what the industry is doing with these technologies. You covered so many, Rajesh, from Edge to RAN to Cloud RAN. What are you most excited to see at MWC this year? And obviously, NVIDIA is going to be there.

Are there any highlights that you're looking forward to from NVIDIA? Yeah, we live in interesting times. Connectivity has become akin to oxygen today, right? It's impossible to even spend a few minutes in today's world without connectivity. And as we discussed earlier, most modern applications will require distributed processing at the edge to meet the latency and quality of service requirements of these applications. And this confluence of 5G, AI, and cloud is super interesting.

And it's, in my mind, game-changing. I think it's really transforming the way we live, quite frankly. So like I explained earlier, to some extent, 5G RAN is a weak link, and it's not keeping pace with AI and cloud computing. And the approach that many VRAN vendors have taken with a fixed-function ASIC-like acceleration to get around the limits of Moore's Law sets us behind in our vision to drive a fully programmable and software-defined infrastructure. What we really need is a general-focus acceleration platform

that can bend the performance curve where Moore's Law cannot. And this is what we are trying to do at NVIDIA. So bring the GPU-accelerated computing and the virtues of standard high volume server platforms to transform the RAN and the edge. I also believe that RAN in the cloud is the future. It's a natural evolution and the next step for the wireless market.

A virtualized RAN built using cloud native technologies is definitely a necessary first step. However, if you can get to realizing the cloud economics for 5G RAN and increase utilization of the RAN infrastructure and to drive the co-innovation of 5G with edge AI applications, we must embrace the principles

of a full RAN in the cloud. This is our focus for NVIDIA's aerial platform, which delivers a fully programmable, scalable, and cloud-native software architecture as a foundational technology for RAN in the cloud. We are actually going to be demoing some of this at the Mobile World Congress in Barcelona next month. You'll hear us talk a lot more about not just AI and cloud computing, but also what we're doing with 5G and, in particular, RAN in the cloud

at NVIDIA's event, the GTC event that we have in March. There's a lot in store over the next couple of months. Just in the broader context, you know me, actually, I'm a dreamer, and I'm truly excited at what the future holds for all of us as we bring 5G, AI, and cloud computing together at the edge to build the applications of tomorrow. Rajesh, it's always lovely to talk to you. And I learn so much every time we do. Thank you for being on the program today. I have one final question for you. Where can folks reach out to learn more about the aerial platform and other things that

NVIDIA is delivering to service the network and edge? And how can they engage with you? You can always connect with me on LinkedIn. And as far as more information about Aerial and what we're doing to transform the radio access network, you can look at developer.nvidia.com slash nvidia-aerial-sdk-early-access-program or simply Google NVIDIA Aerial and it should actually take you to this landing page. Fantastic.

Thanks for being on today. It's a pleasure. Thank you, Alison. It was great talking to you and look forward to connecting with you at MWC and perhaps also at GTC.