AI Networking, Simplified: Inside Cisco’s Nexus Hyperfabric AI - Six Five Virtual Webcast

AI Networking, Simplified: Inside Cisco’s Nexus Hyperfabric AI - Six Five Virtual Webcast

‍

How are organizations overcoming infrastructure roadblocks as they scale AI workloads for real-world production and agility?

‍

Host Matt Kimball is joined by Cisco's Russell Rice, Senior Director, Product Management, for a conversation on how Cisco's Nexus Hyperfabric AI is redefining AI networking through simplified, scalable, and cloud-managed infrastructure. 

‍

Key Takeaways Include:

‍

🔹AI Infrastructure Bottlenecks: While many teams cite infrastructure complexity as a barrier to AI adoption, Cisco’s Hyperfabric AI aims to streamline deployments and eliminate integration challenges.

‍

🔹Hyperfabric AI Differentiators: This solution emphasizes faster, easier scaling of high-performance AI systems compared to traditional approaches, with innovation in management and networking layers.

‍

🔹G200 Switch and Orderability: The newly available G200 switch enables immediate deployment options, accelerating AI project timelines for organizations at both the pilot and the production stage.

‍

🔹Cloud-Managed Operations: Managing both network and infrastructure as a unified, cloud-based platform delivers measurable gains in speed, cost-efficiency, and simplicity for customers.

‍

🔹Evolving IT Mindset: Leaders must shift from custom-built networking to more standardized, modular plug-and-play, AI-ready infrastructure to stay competitive.

‍

Learn more at Cisco.com

‍

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍

Disclaimer: Six Five Virtual Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

Matt Kimball: Welcome to another recording of the 65 virtual podcast. My name is Matt Kimble, vice president and Principal analyst with more insights and strategy, and I am here with Russell Rice, Director of Product management at Cisco. Today we are here to talk about AI networking simplified, taking a look inside Cisco's Nexus Hyper Fabric AI product. Hey, Russell, thanks for joining the podcast with me and thanks for taking a few minutes to have a chat here.

‍

Russell Rice: Love to be here. Thanks for inviting me.

‍

Matt Kimball: Yeah, let's have a fun time. So let me set up the problem statement here. Okay. Kind of get things going and start our conversation. So as we think about AI in the enterprise and kind of moving from pilot or proof of concept to production, we start talking about scale, right? We go from a few GPUs and a few servers to hundreds to Even thousands of GPUs and we go from small workloads to large data management platforms like Vast, trying to feed all of these B300 GPUs from Nvidia and, and it gets really complex trying to connect A to B. And really the most important part of this is that connective tissue that can take all of that data and feed those GPUs quickly. But it's hard, right? I mean, I'm sure being at Cisco, you've heard this a lot from customers. So as you kind of look at what's going on out there in the AI space and you think about companies wanting to really kind of embrace AI at the enterprise level, we know their infrastructure really can't keep up. What are you hearing? I mean, what do you see out there? And, and, and, you know, does. When we talk about, you know, hyper fabric AI, is that really the answer that, you know, enterprise customers seem to be looking for?

‍

Russell Rice: Yeah. No. Good question. I will offer. You know, AI is definitely a challenging thing we see for a lot of organizations to really absorb.

‍

Matt Kimball: Yeah.

‍

Russell Rice: Now, obviously when you hit organizations that are at the hyperscaler and several in the NEO cloud at this point, they've kind of figured out the operational side and how to organize around it and address scale. But as we move into general enterprise, there's a lot of things that need to be addressed and thought through. They're quite different. And I, I'd offer one of the fascinating things about AI is as you consider scale is, it is two vectors that it becomes complex to grapple with. One is just the technological side, right? Hyper fabric is a new offer from Cisco. We're trying to make things simpler here. And even if you just go to kind of the maximum size of an enterprise reference architecture video, which they pump out, which we're compliant with, you know, you're dealing with interconnecting a thousand GPUs, for example, and I don't know if you've stepped back and kind of actually looked at what the infrastructure footprint needs to be for a thousand GPU environment, but you're dealing with scores and scores of switches. You're dealing with front end, back end storage management networks. You're dealing with a server's arrays with that number of GPUs on them. You're dealing with storage infrastructure that deals with large petabytes of data. And even just from a physical wiring perspective, you're probably dealing with over a thousand optical cables, right? So if you just step back and think about what that looks like from a new build perspective, that's very different than let's wire in two servers and go. And understanding what that means early on so that you don't miss your expectations and you make sure that you've got a path from kind of initial absorption to where I want to go is an area that any assistance can help a lot. And that's one of the areas that really, really makes this challenging. I'd say on the other, other, other spectrum that you run into, it's also getting to the parts of the organization that it actually touches. AI is not generally just owned by one team. You know, in order for technology to be absorbed by the enterprise and be really widely used, what you really run into is, is cross marrying, you know, the classic networking term server teams, server operation teams. You've also got AI and AI operations teams. You've got software management teams for software lifecycle management. You also have security that ties into a lot of this AI stuff too. And they're all forming together in this new space they haven't worked together on as well for brownfield deployments. And the other interesting thing that you start getting into here, which is a bit more of a unique vector AI is bringing in, is all of the facilities side because of the power draw and kind of that the real physicality of this thing, it also has an implication to the facilities. And so how can you kind of look at all of that in an efficient way where at the end of the day the speed with which kind of from first decision I want to go down this path to I better be operational to get the advantage and the financial return that I'm looking at for my initial investment out in time before I'm obsolete, which tends to be running on kind of three four year cycles. That compressed stuff makes both of those vectors hard. And so that's, you know, this is different than what we've run before. This is why AI is a bit more challenging than I think things we've had to deal with historically. And so it's kind of exciting and challenging in both regards.

‍

Matt Kimball: It's funny, you know, because I worked in IT in a previous life and when you talked about power budgets back when I was in it, it was almost laughable. It was owned by facilities or operations and it was something IT Net folks never thought of. Right. And, and it's just to what you're saying, it's one of those vectors that really kind of changes the game when you think about AI and you're right, you're right. It's truly kind of an enterprise. Everybody is a stakeholder at this point. But you know, you talk about this and quite frankly I hear hyper fabric sometimes but you know, a lot of folks hear AI infrastructure and fabric and you know, the immediate the mind immediately goes to what the huge cloud providers are doing with you know, all of these GPUs and it's super complex, super expensive in their heads. Right. And it's really difficult to scale. Right. What is it about hyper fabric AI that kind of makes this easier for those IT teams, those folks that are out there and trying to lay out infrastructure to enable these, you know, agentic AI enterprises to really excel.

‍

Russell Rice: You know I think, I think what I'd offer that we took a different view on this than I think is is normal for products in this space and is we step back and try to say okay, let's rethink the goalposts of what we become responsible for as part of our solution design and I mean kind of what we bring to the table and where we start in the process. And so in our early investigations with talking to organizations about absorption, I'll just share with you and I wanted to talk about kind of network and scaling and characteristics like that because hey, I'm Cisco and you know, I'm responsible for the networking, you know, new cloud managed network solution. That's easy. And the first meeting I had was somebody who went back to what I was telling you before. Hey, let's talk about physical cabling. Because if there's 1,000 cables, I need help there. Because my project could take me five months before I've even got everything procured, racked, stacked and up and running. And I've already lost a large part of my budget. And so what we realized is we need to start as a technology provider and think much earlier in the cycle and think through the entire life cycle, what it means for AI to embark upon. So, we did a different tack there. When your first interaction with our solution, you're actually starting at a conceptual design level before you buy anything. You're basically trying to say, what kind of capacity do I need? What type of infrastructure designs does it serve? What kind of GPU array do I actually require? What kind of storage infrastructure do I need? How do I wire and bring all this together? What type of power envelope does this whole thing require for me to do it? What security surroundings do I want around that? And I still haven't bought anything. How can I make that design part of my process and automatically feed into the next pieces of what I need? Like, okay, if I like this, can I make sure all the teams involved, you know, the software team responsible for the AI software, know what they're getting. The team that's doing network management, do they know where they're getting for the network side? The team that's responsible for the servers, do they know what they're getting? Does everybody look at the same design, have an understanding of what they're, what's encompassed? And then can you take the next step? Can you make sure what you've conceptualized and all agreed upon, you actually procure easily, quickly, with all the right parts without an error? You know, having been involved with, you know, passing spreadsheets back and forth with crews, trying to make sure all the optics for a thousand different cables across all of my different wiring diagrams are right. It's really freaking hard, right? You can automate that as part of your life cycle. So come from a design build, bombs build right cabling plans so once you receive kit, how do you automatically just plug it in? It's automatically provisioned the way your design said it should go so that there's no work going from deployment from original design intent. That's the life cycle we focus really heavily on. It's bringing the entire process closer, not just the networking side or can I get enough GPUs but what's the life cycle from concept into operation. And that's where we focused. And it's not that people don't do that today. How can I remove the likelihood of error and the elbow grease effort involved to get there? Can I shrink that whole thing?

‍

Matt Kimball: Yeah, I like that. If you think about a critical path and any project. Right. It's about shortening that critical path and really accelerating that time to the first token which you know, it speaks to outcome focus instead of just, you know, kind of pure technology focus. And I like that approach. It's like you're working with customers to help them achieve kind of where they're trying to get to and stay. And then to me that is really kind of, you know, I think more companies should be kind of focused.

‍

Russell Rice: You know. I'll give you an anecdote. By the way, one interesting thing that happened as part of this. I mean we're very kind of customer centric in the way we built this out. You know, I told you that story about a person saying let's talk about we're on layer one and kind of how do I just build this stuff out? We ended up embarking upon and have realized a whole new part of the what's the helping hands thing that I can give somebody, we call it on site when they're actually got your equipment. Like it's. What's the procedural map for what I need to do to rack it up, to cable it up and to make sure in real time it's right. So you know, we include that as part of the process. And that's not normally what you think about as a network vendor or, or an AI vendor. But it's an important part of this process just to make everything easier.

‍

Matt Kimball: Yeah, yeah, I think that and it's anecdotally I can speak to this as well. Every, every IT person I speak to, whether they're IT administrators, network administrators or CIOs. The common theme is we need help. This thing is really complex and you're right. It's not just that we need help from a network administrator to do a network or a network company to do network related stuff and then a storage company to do storage. We need someone to help us remove all of this friction. Hey, so you know you just made, correct me if I'm wrong, I heard you just made hyper fabric AI Orderable, Is that correct?

‍

Russell Rice: I am very excited to say we made AI Hyper Fabric orderable. It happened this month. Tremendous excitement on our side. I will say, as a proud product owner, I will share with you that we started this project about two and a half years ago. We released Hyper Fabric seven months ago and now we made the AI variant of Orderable. Now that's pretty frickin fast. So we're going startup speed here at Cisco to make this all possible and we're really proud right now just to share as part of orderability, we're actually rounding out our betas right now. So Cisco's finished its own beta, it's rolling production, production live as I speak. I've got several other betas that are rolling out right now that are this week. Momentous stuff like finishing it all up and verifying is going on. So it's dang exciting. I'm expecting and hoping to light this whole thing up within the next four weeks or so, actually three to four weeks and kind of allow anybody to come in and have some fun with this stuff. I think they'll be excited about it.

‍

Matt Kimball: That's awesome. That is awesome. So tie all this together for me. Okay, so we're talking about Hyper Fabric AI. We're talking about, you know, the way you're engaging with customers to, you know, deploy infrastructure and really simplify the process of connecting all of these different points at scale. But now we get it deployed, we're at day two and we're talking operations. Right? You know, or actually let's not even go there. Let's start with, let's stay at day zero first and get to day two. How are you seeing? I think you talked a lot about kind of simplifying the process of deployment, but in terms of like cost, simplicity, kind of, you know, time saved, you know, are you from the early feedback you've received from betas or, you know, kind of, you know, what you've seen in general, you know, are you seeing kind of, you know, a lot of benefit to customers when you talk about network and infrastructure kind of being managed as like this single cloud based system? Or is that just pie in the sky talking? I mean, you know, a lot of, a lot of vendors claim it. Tell me a little bit about what you're seeing.

‍

Russell Rice: Yeah, no, it's a good question. I think, you know, one of the things I'll call out first is yes, we've got some, you know, obviously early but demonstrable feedback about how much this saves. And there's two big accesses for it. Because at the end of the day you use the phrase time to first token. I mean, that's what it's all about. And so every step that reduces, that is important early feedback we're getting right now is the, the amount of time physically involved to get this stuff rolling is like a tenth and order of magnitude difference. And the amount of calendar time expressed. You know, we want to talk about an organization who spent months getting their first AI deployments out with, with our solution, weeks. So that's, that's a different kind of, that's a different unit of measure and we think we can press that. That's the first touch. And so the whole point of that is, is, is speed and velocity. But I think, you know, an underpinning kind of question like how does that material save other than maybe you've just done it before so you can do it faster or something like that. Just skills grow. I think in addition to that lifecycle approach that I had that I was describing before, one of the foundational things that we've done differently and it's really on the networking side that becomes a bit different, is our whole goal of Hyper fabric is to make the customer experience of hyper fabric of the entire network feel as if all you're doing is managing one big switch. And whether or not you've got 10 switches, you've got a complex front end, back end storage management network of a thousand switches. And you know, I don't care because I'm trying to present that as simplistically to you as possible so that you don't have to understand as a customer the interpinnings, nor be responsible for the provisioning and the ongoing operations of what makes that fabric a fabric. Instead you just want to make sure it's providing, it's providing connectivity to your GPUs and it's fast and working. Right? It's providing connectivity, your storage and making sure your storage can talk, you know, to GPUs and it's fast and working. That's what you care about. All that muck in the middle. You know, I want to go to my cloud experience. Cloud ergonomics is generally, I don't understand the plumbing. I just care about my apps working. And what we're, what we've done is we've, we've built that knowledge. That approach in the Hyper fabric analogy I like to give is, you know, AI for people. You don't want to have to build your own car and understand your engine and suspension, all that. What you actually just want to do is drive. And so we've tried to make it so you drive, but that also means that you have to rely on us, that we built the engine that'll work for you, that we've got the suspension that will work for you. And so, you know, there's obviously going to be some constraints around it, but we let you drive.

‍

Matt Kimball: Well, just like, just like buying a, I don't know if you're in my age range, but just like buying a car 30 years ago versus today, you know, you used to worry about, you know, how many cylinders and you worried about all of the different elements of the engine. And today it's a different, totally different buying experience. And that is how technology adoption is, has evolved in the enterprise as well, especially with the, a new demographic kind of coming in and pushing us old guys out. Right. It's an important factor. And I like what you're saying there. You know, a lot of folks reference, you know, that infamous MIT study from a few months ago that, you know, 5% of AI proof of concepts don't, you know, show a measurable, measurable roi. Right. You're talking directly to the heart of what is preventing that ROI from being recognized in a, in a reasonable time frame. Right. So it's a, it's a great story you're, you're telling, but let's, as we're talking through this, let's kind of drive this to the bigger picture. Right? So AI infrastructure is becoming more plug and play thanks to companies like Cisco. And the way you partner with, you know, the Nvidia’s of the world, which is very complex on one side, and the vasts of the world, which is very complex on the other side. You are the easy button or the easy fabric that's tying that all together. You know, so as you think about that and you think about kind of what you hear from IT leaders, IT practitioners, what do you think of, like the mind shift or the mindsets that have to shift a little bit, or the perspectives that have to shift a little bit to help them stay ahead and kind of like really be proactive in driving AI throughout the enterprise instead of reactive to all the problems that are going to kind of inevitably pop up?

‍

Russell Rice: Yeah, no, it's a, you know, I think what we're, what we're really going to be discussing here is kind of what's the evolution of the problem you're trying to solve, you know, from first contact to kind of next steps. And I, I think, I think where a lot of companies on the beginning of their journey that we're talking to are many of them are still at the. Is there a financially viable model that I can, you know, actually operate myself that will work? Right. Is the money in it? Can I do it myself? Honestly, a fair number of people are still at that wedge point. Yeah, I think right beyond that is what we were talking about. Like people who understand that and realize it's possible and feasible and economically viable is interesting. It actually can be compelling for some. The next step is what we were discussing. How quickly can I do this to speed it up, to really, you know, gain a lot of benefit more, you know, and stay on the life cycles fast? Those are the first two steps I would offer . The next big arc that we're going to be running into that you already see in people who are quite mature in this area is it's much more about scaling and agility. Yeah, right. This is about the ability to your earliest point, like, how do I start with a certain investment footprint, whatever it might be, and then grow it where it's working and need to be? And that's where. That's where this becomes important. And part of that gets to be the question of can I reuse what I purchased before and incrementally make change where they're growing or changing some of the componentry rather than rebuild from scratch a new thing which you don't always want to go do, you know, don't burn down one house to build another house, but the other other area that you're. That you get into is. It is not just the reuse of the componentry, but how do I make sure that as my learnings or needs of this AI changes from a business application side, I can reuse the infrastructure that I've got invested in and start sharing it for different business purposes. And it still all seamlessly works well. And that gets to things like security layer and managing and set partitioning and those kinds of things out and DevOps out. Right. How do I start automating this to make it easy for me to do this, whether that's orchestration of workloads and AI loads and all the apparatus around that, which ties in the networking infrastructure as well as the server, that whole thing. Now you start to say how I. How do I apply a tight DevOps model around that? That's where the next kind of areas tend to go. And I think you'll see the hyperscalers were already there. That's what they do. And I think enterprise is going to follow that same footstep.

‍

Matt Kimball: Yeah, I would agree with you and I would say a couple of things I hear out there. I hear a lot of almost irrational enthusiasm from folks around AI and what it's going to do and what it's not going to do. And sometimes it feels like AI is the hammer and everything looks like a nail. It's an old expression. And I think you said this at the outset. I think if you think about outcome focused and what you're trying to achieve to start, that should. And that's at the end. And work left from there, every work backwards from there. Everything should. Should dovetail from that kind of what those desired outcomes are meant to be. And I always, I always, always, always try to adhere to the principle of keeping it simple. Right. The whole engineering, you know, KISS philosophy. Right. Keep it simple, stupid. And it's a. It's easier said than done. And AI is very complex, but when you look at what Cisco is doing with Hyper Fabric AI, that's kind of like applying technology to that principle to really drive that simplicity across the organization. So I love what y' all are doing, and I'm sure it's gonna. It's gonna do amazingly well as you go into production. Before we hang, before we close out, any last thoughts you want to kind of leave us with?

‍

Russell Rice: No, I'm just, you know, for those taking a look at this video, I just offer, you know, this is, you know, I think it's an exciting time for AI. You know, it's definitely. I was working with a compatriot who said, you know, at the end of the day, what we're really trying to do and needs to happen is the democratization of AI. And that really is not just from an individual using it, but the ability to actually host it and manage it yourself. And that's, that's really the journey we're all on right now and how to make it so that's possible. And so it's exciting. And, you know, the phone went through that journey and looked at what we've got right now.

‍

Matt Kimball: That's actually great. It's a great analogy and kind of like the automobile, the phone, it's just, it's a. It's a continual evolution. And yeah, really great words too. To end on. Russell, thank you for taking the time and for those that are out there, please take the time and, and check out what Cisco is doing with Hyper Fabric AI. It's a great story and it really does drive out the complexity and drive down that time to the first token. So thank you for taking the time and hopefully we get to do this again in the near future and talk about the great successes you've had.

‍