Murali Gandluru, Cisco

More from The Networking for AI Summit

Murali Gandluru

Vice President, Product Management, Data Center

Cisco

Info

Murali Gandluru, Cisco

Murali Gandluru

Vice President, Product Management, Data Center Cisco

>> Hello everyone, and welcome to our next Networking for AI Summit session. Today we're fortunate to be joined by Murali Gandharu, Vice President, Product Management for Data Center Networking at Cisco. Welcome, Murali.

>> Thank you, Bob. Thanks for the opportunity. It's wonderful to be here with you and the Cube. I'm really excited about it.

>> Yeah, absolutely. I'm really looking forward to this digging into what's going on in the data center, and it's so important because enterprises are really starting to rapidly adopt AI. Data centers in particular are being pushed to their limits. There's all sorts of new workloads, the large-scale training, inferencing, even the emerging agentic AI applications, there are really demand networks that are faster, more agile, more intelligent than ever before. Yet at the same time, organizations need to make those critical architectural decisions that will impact their ability to scale and compete in this AI era. So I'm looking forward to this conversation as we explore how Cisco sees this evolution of AI networking, what organizations need to consider as they modernize their infrastructure and how these investments can drive both technical and business value. So with that said, let's get started. It's interesting looking over the financials. In the last year, Cisco did over 2 billion worth of AI infrastructure in the last fiscal year, so helping a lot of customers. Can you help us understand a little more specifically how Cisco AI network solutions are helping to shape that future of the data center infrastructure to support these AI workloads at scale?

>> Absolutely. One of the things, like you just pointed out, we are seeing a massive build-out of infrastructure for AI, and fundamentally the network sits at the center of that, because the network sees everything. The other reason why the network sits right there in the middle is because there's a lot more of the distributed infrastructure going on, where you have moving from a training-based environment to a inferencing solution. Cisco's approach to AI networking takes on whatever we have been doing from the beginning of the data center switching paradigm itself, which is silicon, differentiated silicon, systems, software optics, the complete integrated approach. It's more essential than ever because we are delivering AI infrastructure at scale, not just from the perspective of bandwidth, but also from the perspective of telemetry, the ability to provide deterministic behavior across a variety of complex workloads. We are no longer just talking about just GPU-to-GPU communication, or GPU-to-storage. You have hybrid environments as well. So effectively, this three things. One, massive east-west bandwidth, sometimes spanning across sites. We have to break through the leaf spine architectures that we are used to provide a different kinds of load balancing, provide deterministic, low jitter transport and life cycle tooling. That's an important piece, the life cycle tooling and management. Finally, the third one, what I would say is built-in observability, automation and also the ability to be able to quickly pinpoint issues when they occur and be able to take those gracefully out of service, and then bring them back gracefully in as well. So this is how we plan to address the architectural shifts that are happening in the network thanks to AI and the networking paradigm itself shifting.

>> Yeah, excellent. I think that's great, and I think you brought up something that's super important for organizations to understand. I think up until now, a lot of organizations have been talking about how they scale up and scale out those AI data centers, but you even brought up that scale across concept as well, in going between data centers now, especially as organizations try to find sources of power and so forth, renewable power and whatnot. So it sounds like that's going to be something organizations really need to consider as they're planning these environments.

>> Absolutely, they are. I mean, scale wide maybe then, in this case across sites. The networks are primarily designed in such a way earlier where you have scoped domains, let's call them. When you have a hyper distributed enterprise with complex operating environments due to communication paradigms, shifts, essentially, you've got to make sure that the organizations managing this, monitoring this and also scaling them are able to continue to be able to do so at a significant clip. So fundamentally, what we have to do is that the network sits at a great place where it's not a passive transport layer. We have to also incorporate elements of security. Security is so important and has to be infused into the network, and that's the theme that we have been going on and talking about to all our customers, and bringing products to the market where we infuse security into the network as well, bring policy enforcement which, more than ever, as you know from our software defined networking paradigm, we've had to do it, but more than ever enforce policy, centralized, distributed, I'm sorry, design policies centrally and distribute it and force it across your infrastructure. And that's where the smart switch comes into picture as well. So security, AI, automation, and the network, it sees everything.

>> Yeah, I think that's super important because in these AI data centers, there's always been historically a trade-off between performance and security. And the reality is today there is no more trade-off. It has to be both highly performant and it has to be secure. So I like your concept of that centralized policies and distributed enforcement is what's really going to help organizations, and especially in these highly distributed and all the East-West traffic, you need to be able to impose those security policies at every point along the way, and so that distributed enforcement is going to be super important. There's a lot of enterprise networking teams out there. They've been working in the data center for a long time, but the AI environments are really very different from prior data center networks. So what can you share with our listeners now about the unique demands for AI training, inferencing, either agentic workloads, placing on that data center as compared to what they're used to?

>> Yeah, I think one of the amazing things about ethernet is that over the past 30, 40 years regardless of the technology, ethernet has evolved to become that democratization mechanism, and democratization of infrastructure will happen, is happening already. And the reason for that is because ethernet has been built in such a way, architected in such a way that it provides ubiquity. First of all, there's ubiquity of deployments. The operating model of ethernet is everywhere. Across all infrastructure now, not just the data center infrastructure. Then there's the ability to drive that ubiquity with the right kind of economics. That is a key piece, whether it is switching, NICs, cabling, optics, that ties into that cabling, all of them have essentially been built around ethernet for, I don't know, about a good 25, 30 years. And the amazing thing about ethernet is that it's evolved across varying feature sets, speeds. I remember we started from a hundred meg to the axis, all the way now where we are talking about a hundred gig to the axis, 400 gig to the axis, we have 800 gig clusters and beyond. Support for capabilities that Cisco has had for a while because of its fidelity for storage traffic, we have been working on RDMA, Rocky V-II capabilities that have been there in ethernet for a long time. So it can deliver low latency, high bandwidth characteristics that you need, and it can evolve across all of these new innovations, including things like new data processing units like the DPUs or SmartNICs. You can mix in CPU, GPUs, like I just said, in a network. All of that, I also want to point out, is Cisco has always taken a complete end-to-end picture, which means that the ability to drive an operating model consistency, for example, the Nexus in the data center and networking piece, allows us to be able to have a common tool chain, common skill set, and a common approach to managing infrastructure and then farming out data from that infrastructure in a consistent way. All of that reduces operational risk and improves the TCO model as well. So in a simple way, what I'd say is that ethernet gives customers the performance needed for modern AI workloads while managing existing non-AI workloads, preserving this paradigm of simplicity, scalability, seamlessness in the existing operations, and of course all of this in a secure approach as well.

>> Yeah, no, I think you're on track there. I remember a friend of mine went to a Bob Metcalf talk and he said, "I don't know what's going to come after Ethernet, but I'm pretty sure it's going to be called Ethernet." We've seen that over and over, and research that I've done actually showed almost 60%, 59% of organizations, enterprise organizations said they'd prefer to use Ethernet over InfiniBand, and for a lot of those reasons you were highlighting. They've got the skill sets, they use it everywhere else. So their ability to scale for organizations that were going within InfiniBand, it was largely because it came bundled as part of the package.

>> Exactly.

>> So I think organizations are recognizing that in order for this to scale effectively and manage it, they really need to have Ethernet across all parts of their environment. And I know you're part of the UEC, the Ultra Ethernet Consortium and so forth, all those things that are helping to make sure that Ethernet will be able to support all these AI workloads.

>> Absolutely, absolutely. And you're right, that ultimately the next big thing that's coming will be tied to Ethernet in some shape or form as well for both scale up, scale out and, as we said, scale wide as well.

>> Yeah, all the new dimensions that have to be taken into consideration with this. And it is great to see, like I said, Ethernet evolving to handle these new tasks that are being assigned to them as well, if you will. It just shows the brilliance of the technology, to be able to adapt and morph to what's needed. Wanted to ask you a little bit, again, we've got a lot of people out there getting started, trying to figure out what they should do. So I'm hoping given, as I talked about when we started, the experience that Cisco has now, you've got a lot of enterprise customers, a lot of hyperscalers and so forth behind you. What are some of the best practices that you would recommend for an organization who's trying to build out that agile future-ready data centers to accommodate AI? And also as AI shifts to the edge, how does that also impact what an organization might want to do?

>> Yeah, I think best practices, like in any other time of disruption is very critical. How you design your infrastructure and how you plan for that design, as well, has never been as important as now. So what I would say is for today, build on a unified Ethernet fabric, and so build and expand from where you are today. Think about how you can deliver some of these elements. The non-AI elements already exist, but think for your infrastructure to be able to support that existing model and paradigm while you go onto the next stage as well. And in fact, that's one of the reasons why Cisco has built out its own differentiated silicon portfolio with SiliconOne, which allows us to take care and make sure that customers are able to seamlessly evolve, but that's one. Second thing is, build out in such a way that you are able to build blocks of modular nodes or fabrics, and then bring them together using a user experience and an operating model that you are very familiar with that allows you to scale across multiple environments. And those are things that actually will help you set yourself up for the next stage. See, I'm mostly always thinking about the operational complexity of new blocks coming in. Whether it's front-end, back-end management, storage networks, you want to ultimately get to a common operating model, and that's very important. So I would say looking ahead, think about a platform mindset, and this platform mindset increasingly will not just be about the network, it'll also be about the security that's infused into the network, whether it's about protecting the models that you might have in distributed environments, whether it's about protecting your workloads as they communicate with the storage clusters, whether it's about just making sure that there's a fairness in your fabric itself to be able to load balance traditional workloads, and also spray packets at the same time, while being flow aware. So all of this means that you have to have significant amount and also that it transcends across hybrid environments. So observability, being able to provide nice telemetry, these are the key.

>> Yeah. I think that makes a lot of sense. I also wanted to touch upon, again, thinking about our audience. They're trying to get started. Going with ethernet is sort of the next logical step, but what else are you doing to help accelerate their journey? If I remember correctly, you've got a lot of validated designs, Cisco validated designs and so forth. How should organizations think about getting started deploying their AI data centers?

>> Yes. Fundamentally, I think Cisco's approach has always been about building validated designs, white papers and other kinds of collateral that aligns with verticals, either segment verticals or, for that matter, any kinds of different model verticals as well. So the validated designs is the place where it starts for an architecture. You tie it to white papers and other kinds of videos and content that we provide that allows you to understand what is impactful, for example, in financial services. So we know that financial services, they have an immediate use for AI already. How do you make sure that that AI usage is also tied to make sure that it's secure, there's no fraud, and how do you tie it to other tools? So it means its solution mindset is important. So financial services, they'll have their own healthcare environments, autonomous vehicle environments, manufacturing. The industries are the same. It's just that they will have to adapt. And so that's where Cisco makes sure that your business outcomes are what are important, and that this network is purpose built for AI and it evolves from where you are today. So we are there across the industries, and we communicate through these various validated designs and papers and so on to be there to help whenever needed. We, of course, would love to hear from you. We want to make sure that the value that you get for the investments you put and the change that's coming, you're seeing the best out of that. So that value, it starts with the business, then it goes into the human operator elements as well. So we want to make sure that the business impact and the impact of the network, broadly, for AI is there for you.

>> No, I think that's a great approach, and I think especially now as organizations are getting started and by vertical being able to help them with the use cases, and this is how organizations are doing it, will certainly help to accelerate it. And that really ties into one of my last questions that I have for you, is that when we think about deploying AI, there's been a lot of talk in the industry about where's the ROI, are people gaining benefit from it? I know from attending events like Cisco Live and so forth that your use of AI internally is enabling you to really accelerate the innovation train and get going faster. I'm wondering if, with the experience that you've had, if you could also talk a little bit about how organizations who are deploying it today are starting to get measurable business value, and maybe how IT leaders should be thinking about framing those investments to their board or business stakeholders about, "This is why we want to do this."

>> Yeah. So we just had our customer advisory board in New York where we spent some time on this topic, about how to figure out what their existing projects are that will be tied to the company's AI strategy that derives the maximum benefit for that particular vertical, or for that particular segment. And essentially, we think the organizations and the infrastructure team is critical. In fact, has never been as important as today for providing insights and the value that you get at, first and foremost, compliance and governance. How AI is helping the compliance and governance elements, because you've built infrastructure over the past 10, 15 years that has a way for measuring governance and compliance and audit reports and so on. Now with AI, that is dramatically simplified and hopefully, you're able to provide insights to your CIO in a much more real time and beneficial way. We are certainly seeing that internally in our organization. Cisco IT, who was also there at our customer advisory board was saying, "Look, we are using all the differentiated technology that Cisco builds, especially now with infrastructure for AI and, very importantly, AI for infrastructure." So things like our Nexus dashboard with the ability to do job monitoring in real time basis, and the ability to actually scale across multiple sites, multiple locations, provide real time insights and orchestration of nodes, as well as narrowly and gracefully taking out services. These are things that will show up in reports to your CIO, and you should show it, how you're able to do it so much faster. So that percentage, how much time you've saved, both from an operational complexity perspective as well as from a number of resources perspective, those are things that will definitely give an immediate value. And so we continue to see that things like our AI Canvas that was announced at Cisco live, our AI assistance for networks and, for that matter, expanding it into compute and to the NICs and other elements of the GPU, those are things that speed up how we can benefit from this infrastructure. So these are things I would say.

>> Excellent. Yeah, I think those are some great examples. And yeah, I know I've been a big fan of the multiplayer capabilities of AI Canvas being able to tie in, again, when you were talking earlier, networking, security, other teams, observability apps, teams, et cetera. So certainly a lot of capability for organizations to take advantage of, so that they can accelerate achieving their value from their AI deployments.

>> Absolutely. It's an exciting time. It's a really exciting time to be here.

>> Yeah, absolutely. It is interesting. Some of that research I had done last year, 93% of the respondents came back and said the network is more important to achieving business goals than ever before. And so that's a pretty huge amount, right? Everyone understands now the network is back. Super important, certainly for accelerating these AI deployments. This has been an awesome conversation, Murali. I really appreciate it. Thank you for sharing your insights.

>> Thank you, Bob, and thank you the Cube for the opportunity.

>> You got it. So we heard today about the critical role Ethernet's going to play in scaling the infrastructure, a lot of the architectural shifts that are going to be needed to support the inferencing and agentic AI workloads, and a lot of best practices as well for building out these agile, future-ready data centers that connect seamlessly from the edge to the core. So I think for our audience, one of the key takeaways that's clear is that AI is really redefining the data center, and networking will be that foundation that enables organizations to scale, innovate, and deliver business value. So thank you to all of you who are joining us today in this very important discussion. Stay tuned for more Networking for AI Summit.

Murali Gandluru, Cisco

>> Thank you, Bob. Thanks for the opportunity. It's wonderful to be here with you and the Cube. I'm really excited about it.

>> Exactly.

>> Absolutely. It's an exciting time. It's a really exciting time to be here.

>> Thank you, Bob, and thank you the Cube for the opportunity.