Thanks for joining us today at the Raymond James Tech and Consumer Conference here in New York. My name is Simon Leopold, Raymond James Data Infrastructure Analyst, and I'm pleased to introduce Rakesh Chopra, who is from Cisco and is a Cisco fellow on the Silicon One team. Structure for this presentation is roughly 15-20 minutes of prepared remarks, help folks learn about this business, this part of where it fits into Cisco, and then I've got a few questions prepared, and we'll give the audience a chance as well. We're also joined by Sami Badri, who recently joined Cisco, leading their IR efforts. With that, let me hand the floor to Rakesh.
Thanks so much, Simon. Nice to meet everybody. I'm looking forward to talking today. I'm gonna focus today talking about Cisco Silicon One, its value proposition, what we're trying to do with Silicon One, and talk about how it's applicable... or results are different materially from those in the projections or other forward-looking statements. So with that out of the way, let's jump into it. So I wanna start today talking about where we are from a silicon industry from a networking perspective today. If you go out and you look around, I would contend there's about 20 different silicon architectures that exist in the industry. Now, if you go in and you talk to those various different vendors who are building them, they'll tell you that they need to exist for various reasons.
Some people will claim that you need to have a unique architecture for routing versus switching, deep buffered versus programmable, different scale metrics. A whole bunch of different rationales will be used to sort of explain why there's different silicon architectures for routing and switching perspectives. But obviously, there's also another piece of that, and that's the business side of it. Who's building the silicon? Is it a system vendor, or is it a third-party silicon vendor? And there might be multiple companies going after the same market as well. As these companies continue to invest in these silicon architectures, you create what I like to call technical debt, i.e., it is a thing bringing in the dollars. It's very hard to stop spending money on things bringing the dollars to do something new. Now, what's the effect of all of these different architectures?
We believed as an industry that you can hide this under software for a very long time, but I think you can only hide a certain amount of it. If you think about what this means to somebody who builds equipment, like somebody like Cisco or a web-scale customer who builds their own equipment, basically, at the end of the day, it means duplicated development. You're spending all of your time replicating very similar products under different silicon architectures over and over and over again, decreasing your pace of innovation. If you think about somebody who's buying equipment built by somebody else, the effects are similar, but slightly differently. They have to spend a long time qualifying all of these different silicon architectures. It creates a very fragile infrastructure because you end up working around all sorts of nuances that exist in the silicon architecture.
And again, the effect is the same. You're spending all of this time doing work that is not particularly adding value to your end customers, as opposed to spending your precious dollars on what matters the most. So what we realized with Cisco Silicon One, and at this point, it seems like a very obvious statement, but at the time, it was quite innovative to us, that at the end of the day, if you approach the problem with the same organization and the same technology, you're gonna get the same outcome. And that's what we've been doing at Cisco for a very long time, and frankly, what all the rest of the folks in the industry are still doing today. So what did we do with Cisco Silicon One? We started off by building a new organization.
So we have one organization that's designed to build one architecture to be deployed across the network and across business models. Or said another way, we just set our scope correctly. The second thing is, there's real technology barriers which stopped us and others from sort of accomplishing this task, and we spent years investing in technology at the low level of the silicon pieces to actually sort of allow these things to converge. And finally, we dumped Cisco dollars on top of it. We've invested over $1 billion and over eight years of time behind Silicon One. The culmination of these three things, I would contend, is actually nothing short than a fundamental change in the industry, and we've created what I like to believe, and you will see over time here, is the industry's first truly scalable networking silicon architecture.
We are playing with a very different set of cards than everybody else in the industry. So our goal with Cisco Silicon One is not to be the twenty-first architecture. We wanna be the one architecture for the network. Now, we realize that convergence isn't enough, and so we believe that at the end of the day, we have to take this architecture and build many purpose-built devices. So requirements are gonna drive unique architectures. They're gonna drive unique devices from Silicon One. We've grown up enough to now offer our technologies to our customers in multiple different business models: silicon only, white box, or the traditional full system model. And as we've talked about, we've invested a lot in this in sort of wiping away many of the technology barriers, and at the end of the day, we believe this translates to a better and easier-to-maintain network.
Now, behind all of this, we sort of came to the realization that if you look at what's really limiting the industry today, it's actually power. It is the fundamental limit in networking and compute, and if anybody tells you something differently, they haven't actually quite grasped what is really happening in the industry. Now, why do I make that statement? I make that statement because if you look at what limits what we can build inside of Cisco at this point, inside of the systems or the silicon, it is power. How much power can we deliver to the chips? How much can we cool? If you look at what our customers are seeing, they're out of power in their facilities, in their rooms, in their racks. It's limiting what they can deploy.
We all know that there's a big green initiative going on as well in the world. Our planet only has so many resources. So the way I like to think about this is it's created, for the first time, a technology imperative, a business imperative, and a moral imperative to solve this problem. This is a great catalyst for us as engineers to sort of focus on the problem, and that's what we've done with Silicon One. Our belief is, at the end of the day, that we can deliver sustainability through technology, and you'll see this sort of across Silicon One's offerings in terms of focusing on power efficiency. Now, when we think about the innovation strategy, at the end of the day, it breaks down into a few different pieces. First is you obviously have to have differentiated products in the industry.
The customers have to want to buy what you have, okay? The fact that we managed to converge all of these different markets allows us to drive great leverage across our portfolio, minimizing our development costs and maximizing quality and time to market. Third is we've built silicon for a long time at Cisco. We're building it very differently today. We are a true fabless semiconductor inside of Cisco, and what that does is that drives our costs way down from everybody else from a systems perspective and allows us to offer our technology to third parties. We then take that innovation, and we go and we win customers. As we've mentioned before, our goal is to meet the customers where they are, so we offer silicon only, white box, and full systems, okay? Not every customer wants one of these things.
We're the only company in the world who offers all three of these simultaneously, and that allows our customers to sort of decide where they want to consume our technology and how they want to consume that technology. We've focused a lot on trying to reengage and win the web scale business. We've had great results on that. We'll talk about that in a bit. But the reason to win web scale is they're a great customer base, they're very large customers, but they also drive technology innovations. So winning those customers bases gives you the long-term view of where the industry is going and making sure that all the rest of our business at Cisco also has leading technology. Because they buy a huge amount of components, that also drives revenue in a very healthy way for us at Cisco. Now, as the volume goes up, our scale increases.
That allows us to drive our cost down, and that allows us to then take those dollars that we take and reinvest it back into this machine, and it creates a cyclical feedback loop. Now, over time, Cisco Silicon One has been doing a lot of different things. We announced back in December of 2019. Over time, roughly about twice a year, we come back to the market to talk about more things associated with Silicon One. We have been expanding from where we started in the core service provider peering and DCI roles into web scale switching, into enterprise campus, into the enterprise data center, DCI and spine roles, and finally into AI/ML, moving from 400 gig to 800 gig. We are moving at a pace that is about 11 x faster than anybody else in the industry.
The proof is in the pudding in terms of what I've just said, in terms of how fast we're moving. At this point, again, we have very broad coverage of markets that we're covering with Silicon One. It has been an exciting few years for us at Cisco and I think for the market. And again, to reiterate, our goal here is multiple devices of all trace to one architecture, trying to give our customers both convergence and best of breed. That's really, really important. One is not enough to sort of change the industry as a whole. Now, how have we been doing? At this point, if you look around at various different Cisco products that have absorbed Silicon One, we have the Cisco 8000 running IOS XR, targeting sort of web scale and service provider based applications.
We have the IOS Catalyst 9500X and 9600X running IOS XE, targeting enterprise campus and enterprise edge. We have the Nexus platform, the 9800 and the 9232E, running NX-OS, targeting enterprise data center. We also offer, as we've mentioned, our silicon in non-traditional fashions. So there's a bunch of people out there in the industry building their own hardware, bringing their own OSes on top of that infrastructure. Now, another way to think about how we've been doing is how much have we managed to penetrate web-scale. That is the hardest customer base to enter, and frankly, if I'm honest, we as Cisco missed that transition many, many years ago.
Silicon One is our ability to get back into the market, and if you'd asked me four years ago where we'd be at this point, it would be nothing like this. So we have managed to capture five of the six major tier one web-scale customers with Silicon One, using silicon only or white box business models. So in the data center, non-traditional business models, we've penetrated nearly all of the major customers in the world, okay? That is a huge accomplishment for Cisco and something that I think we're very proud of. I mentioned power before. It's great for me to talk about power, but it's much better when our customers actually run their own experiments. So there's a bunch of different press releases out from customers. I've picked one here.
Deutsche Telekom has run an analysis, and by adopting Cisco Silicon One and the Cisco 8000, they've dropped their power bill by 92%. So this is not 1% or 2% that we're talking about. This is a seismic shift in efficiency. It's part of the reason we're so excited about Silicon One. Everything that I just talked about Silicon One is actually interestingly also applicable to AI, so let's talk about that for a little bit. But before getting into the Silicon One piece, I think it's important for us to realize that it is very confusing to think about AI and Cisco. Cisco is a huge company. We have many different products, many different business lines, and so when you talk about AI, it's quite difficult to understand what does Cisco and AI really mean?...
In my simplistic mind, I like to break it down into two. There's Cisco using AI to improve the products that we offer to our customers. So for example, Webex noise canceling, or Wi-Fi performance, or improved network automation. All of this infrastructure can use AI to offer products to our end customers, which are better. The other is actually selling Cisco products to enable our customers to build AI networks. So this is what I'm here to talk about today. The other is incredibly important as well. I'm just more focused on the hardware side of the equation, which is the right side here. So if I think about the web-scale data center, it is a very big network.
We tend to focus on the top side of this picture, which is the front-end network, which is a bunch of computes connected to each other and to the outside world. This has been a Ethernet network for a very long time, and this is the network that Cisco has been playing in, offering both switching and routing platforms, selling optics systems and silicon to each one of those. This is the traditional web-scale market that we all think of. But there's this entire back-end network that is now becoming incredibly relevant for everybody now that everyone is out there talking about AI. I will contend, and I'll explain why in a second, this network is moving to Ethernet.
It has historically been InfiniBand, it will be Ethernet in the future, and this opens a brand-new market opportunity for Cisco, for us to sell silicon systems and optics. So again, this is quite an exciting time for us here at Cisco as a consequence. Now, why Ethernet? So if you look out today, there's obviously a single company out there who builds the GPUs that power most of this stuff. It was built for high-performance compute, using a sole source GPU technology, sole source switching technology, and an InfiniBand as an interconnect. As AI ML has come out, this market is absolutely huge, and everybody is building GPUs. You see it from Intel, you see it from AMD, you see end customers like Meta, Google, Amazon, Microsoft, all of them are announcing their own GPUs, okay?
As they do that, they have to decide what switching technologies are they gonna use, and they want a broad ecosystem, so they're going out to multi-source switches. The sum of those two things that I just mentioned there translates to the only viable answer is Ethernet, and you can see this migration happening now in the industry. It is not the majority thing that's happening immediately, but it is where the industry is going. So now that we've talked about the fact that it is going to be Ethernet, that opens Cisco to play into that market. So the key question is, why would somebody select Cisco rather than somebody else? At the end of the day, we have the industry-leading silicon with Cisco Silicon One, we have amazing systems, and we have amazing optics that we sell as a component as well.
Again, we're uniquely positioned that we have all of this technology in-house inside of Cisco to work with our customers for what they want, okay? I've also mentioned before that we are now meeting our customers where they wanna be met. Some customers wanna buy full systems, some wanna buy components, some want Cisco to build hardware for them, and they port their OS on top of them. All of those are possible to our customers. That gives us a very broad net for us to be able to engage and talk with our customers and talk about trade-offs with them, so that they can optimize their final business.
Now, if I talk a little bit about the silicon piece itself and why Silicon One is unique compared to other Ethernet switches in the market, it's because at the end of the day, everybody else is playing a one-horse race. We, with Cisco Silicon One, are playing a three-horse race, right? And every horse has its pros and cons. Other customers or other vendors will go out and try and sell the reason that their horse is the best. We'll work with the customers to explain what's good about each horse, for them to sort of analyze what's important for them. We can run their devices for Ethernet, with standard Ethernet. That is the most generic solution, giving you the ultimate compatibility.
On the far end of that spectrum, we offer our customers a scheduled Ethernet interconnect that gives you the ultimate performance, where GPUs are being the most efficient they possibly can. And then we're taking pieces of that technology and layering it on top of Ethernet, or what we call enhanced Ethernet, to pick up some of those characteristics, but still take the sort of wide ecosystem play with Ethernet. And this is really where Ultra Ethernet Consortium sort of comes into play, okay? Now, that wouldn't be interesting if the difference in performance wasn't huge. If we compare the performance of Ethernet to the performance of a scheduled fabric, we can shrink the job completion time by a factor of two. What does that mean?
For half the network, you can get the same performance on your GPUs, or you can put twice as many GPUs on the same network. Again, going back to that notion of sustainability, if you want to move the needle, you have to offer a product that really, really changes the game in a very fundamental way, and that is what Cisco Silicon One does. Now, what I just talked about there was sort of talking about optimizing the job completion time, how GPUs run their AI networks. But Cisco Silicon One is interesting in many other ways when we take that sustainability angle. One is that we've managed to double our performance. Lots of people have done that as they move from 25.6 to 51.2, but we've also halved the latency of our chips. Nobody else has done that in the industry.
We've also doubled the power efficiency, i.e., our 51.2 is the same power as our 25.6. Nobody else has been able to do that in the industry. And finally, this one sounds a little bit, engineering-specific, shall we say, but we're also unique in the fact that we have done our 51.2 with a full 512 radix. That means we can connect to 512 different things simultaneously, okay? That doesn't sound like a big deal, but everyone else has been throwing things overboard to try and fit their designs in. What that translates to when you look at building an AI ML back-end network, is that to connect 32,408 GPUs, we can do it with one less networking tier.
which saves 50% of the optics, 40% of the switches, and obviously one-third of the network tiers. That translates to 1 MW of power savings for a single 32K GPU cluster. Again, sustainability and power efficiency is a huge deal for us here at Cisco Silicon One, and we're seeing good traction with our customers as a consequence. So why do we win at Cisco Silicon One? It's because we have the right technology, we have the right investments, we have the right scale of investment, we have the right cost points, and we have the right business models. We are also the only company who's got all of the technology across silicon, hardware, optics, and software, and that is allowing us to innovate across these technology boundaries. And we're the only company with all of the business models: silicon-only, white box, and full system.
We are meeting our customers where they want to be met, and offering them solutions to meet their real problems. With that, I've reached the end. Happy to take any questions, Simon.
You nailed it. Right, right on time. Yeah, I've got a handful of questions, and we'll check with the audience as well. One of the things that I'm interested in hearing about is how the business arrangements are made inside Cisco, and let me explain what I'm getting at here is, there are teams responsible for making your switches and your routers, and they may wanna choose different silicon, or their customer may want different silicon.
Yep.
How does the... What's the inside working of when do they use Silicon One? What's their incentive?
Mm.
What if they don't want to? Do they sort of get in trouble?
Sure.
How does that work?
Yeah, it's a great question. So, several pieces to the story. So the way I like to think about what Silicon One is, is at the end of the day, you can think about Silicon One as basically being a standalone fabless semiconductor within Cisco. Our goal as a Silicon One organization is to make the best products that we can, that we think are differentiated and interesting to the market, right? That is our goal, and we enable our internal customers, and we enable our external customers. For us, they are all customers. So if you think about it from the other angle, regular Cisco is building systems based on Silicon One or other devices in the market. They're completely free to choose that. And we have plenty of customers who want different choices, and we have products built by Cisco on other pieces of silicon.
We only invest in the silicon that we think we can drive true innovation and true differentiation in the market. It doesn't make sense for us to drive something that looks a lot like somebody else's piece of silicon, right? There's no real value at the end of the day. And so what you're seeing is you're seeing Silicon One sort of extend from the Cisco 8000 to the Catalyst 9500X, 9600X, and Nexus 9232E, and 9800, is those internal customers selecting Silicon One as their offering of choice. Does that make sense?
It does. It does. And I think one of the things I've sensed investors don't fully appreciate is Silicon One is not one thing.
Correct.
It's a family. What is the breadth of the portfolio?
It's a great point. I think when we first announced Silicon One in December of 2019, I think what most people took away was that what Cisco was proposing is we're making one piece of silicon for the entire network, and that was a 10.8 Tb device, deep buffered routing device at 10.8 Tb. If you think about that just for a second in hindsight, now that we've announced more, that clearly doesn't actually make sense. There's tons of places in the market where you don't even want 10.8 Tb of throughput, and so you're not gonna be competitive if you try and do that. What you're seeing play out over time is exactly what you've just said, which is, over time, we're making more and more devices with Silicon One and expanding our footprint. Okay?
So, right now we've got 14 different devices all under the Silicon One umbrella. Again, that is an insane pace of innovation, okay? Again, typically, a silicon architecture spits out a piece of silicon once every 18 months to three years. The fact that we have 13 devices in 40-some months, 42 months, just shows you how quickly we're able to iterate on that. So at this point, we cover basically the entire AI back end, the entire web scale data center front end, toward leaf spine DCI. We cover the service provider core and peering cell roles, and we cover the enterprise core and edge roles.
So I wanted to follow up with you on the power savings issue. So preaching to the choir, I get it.
Yep.
Totally with you on it, and I don't think investors do appreciate it sufficiently either.
Yeah.
I guess when I've sort of thought about what are the key factors, I've felt that when we're talking about silicon, it has a lot to do with process node.
Yep.
That if you go from 7-nanometer to 3-nanometer, you consume less electricity. So in the Deutsche Telekom example, was that a comparison to an older technology and your just new process? And the crux of the question is, what if another silicon provider is using the same process node-
Yeah
... do you still have an advantage? What gives you that advantage?
Yeah. So excellent question. So, let's sort of peel back the onion a little bit. You're absolutely right that, that process node is our friend when it comes to power efficiency. Generation on generation of process node, the power efficiency increases, or the power comes down for the same bits that you push through. That is mostly in the core, the center of the silicon, not the I/O, not all of the SerDes that, that live around it, but that's absolutely on our side in terms of power efficiency. But that doesn't explain the Silicon One advantage. If we were to compare like for like process node for any other silicon on the market, nobody is even close to what we're talking about with Silicon One. I'll give you an example. If you look out in sort of the deep buffered routing space, for example, okay?
There's recent announcements about sampling of a 14.4 Tb piece of routing silicon. They call it 28.8 Tb because they're playing half duplex math to sort of, spray paint, shall we say, some, some, marketing math around to make it look like they're not far behind. That is a 5 nanometer device, okay? That is the first time since we announced Silicon One back in December of 2019 in 16 nanometer, okay, that somebody has higher bandwidth. Okay? Now, why is that important? Within a process node, pretty much all of these devices are the biggest piece of silicon that you can buy, right? These are very complex designs. So if somebody is able to get significantly more bandwidth out of that same piece of silicon die, that means it's inherently cheaper, it's inherently more power efficient, okay?
Nobody else is really playing the same game that we have with Silicon One, and it goes back to some of the technology that we invented as part of the innovation in Silicon One. If you look at the physical block diagrams of a Silicon One device, and you compare it to anybody else's silicon out there, it looks very different, okay? We don't move packets around inside of our devices. They just sit in a memory, and we're manipulating the descriptors. Why does that matter? It matters because moving data takes power. If you don't move the data, you don't take the power, and we've really innovated in terms of memory structures and enabled to enable convergence of routing and switching and power efficiency. Does that make sense?
Somewhat.
Okay. So you wanna ask a follow-up?
No, I'll get myself in too deep at that point. I do wanna ask a little bit about the transition to Ethernet from InfiniBand.
Yeah.
So it's been sort of a hot topic, clearly, and you've got sort of an added differentiated view with Scheduled Fabric, but do you have predictions or thoughts over the timeline for that transition?
Yep. So I think there's analysts out there who've done their own predictions, which we've probably all read and all seen, which I think they're very good projections. The way I think about it now is there's already large-scale deployments of Ethernet AI-based fabrics in web-scale infrastructure. So it has already happened in production. There's other customers who are actively trialing Ethernet AI-based infrastructure. Many of the GPUs that are being built are supporting Ethernet natively or Ethernet-based NICs, if it's not a native integration. And even the incumbent, who let's just say has a vested interest in terms of InfiniBand, is also building Ethernet-based AI infrastructure. The transition is inevitable. It has already started, it's in production. It is...
will probably play out over the next few years, is, is my guess, where Ethernet becomes a sort of majority stakeholder in terms of Ethernet backend interconnect.
Okay, let me check with the audience, see if anybody's got questions. I've got more if you don't, but-
Yeah.
If you're up front.
So-
I'll repeat the question for the webcast.
Okay. So my name is Kishan Sridhar, and I'm from Raymond James Technology, actually from the innovation team. Just curious, when you talk about meeting your customers and Silicon One, are you building anything perhaps like in the sentiment analysis space using Silicon One? What type of work are you doing in that area? Just curious.
So I just wanna make sure we're using the same terminology. So do you wanna define your, your definition of sentiment analysis?
Sentiment analysis? I would call it, technology or maybe even intra infrastructure that we would hypothetically deploy to, as you said, get an outcome-
Mm-hmm
... tied to potentially a client and something.
Yep.
But obviously we're simply exploring what's out there. We are not doing anything in that space or formally researching that.
Yep.
Just wanted to get a sense from Cisco on what you're doing in that space.
Yeah.
Kishan, could you paraphrase the question for the webcast? Just-
Sure. So, I think the question is: what is Cisco? Please correct me if I'm wrong.
Yeah.
What is Cisco doing in terms of enabling our customers to build infrastructure to model or understand outcome of some sort of transactions? That's the basic of the question? Yeah, so-
Talking about particularly in sentiment analysis.
Particularly in sentiment analysis.
Mm-hmm.
So, the way I sort of think about this is, so my area of expertise is what I'll call the silicon piece of the puzzle, i.e., us building the right silicon for our customers. We then sell that to internal and external customers, as I mentioned. There's a bunch of... I can't go into the details of what our external customers are doing, but you can imagine that our external customers are building systems and on around Silicon One with software pieces on top to be able to run all sorts of behavioral analysis style of situations. But that's for others to sort of comment on.
In terms of internal customers, the answer is a little bit the same, which is we have a bunch of Cisco products being built around Silicon One from a hardware, software, and network orchestration piece, but it's really for those business units to comment on what they may or may not be doing in this space.
So we are just out of time. I wanna ask you just one question in closing.
Sure.
Sort of the opportunity of what do you think the investment community doesn't fully appreciate about Silicon One that you wanna make sure they do?
Yep. So I think at the end of the day, I think it's quite easy to maybe miss how fundamentally important something like Cisco Silicon One is, right? We have a tendency to all focus on end outcomes, which are, of course, incredibly important. But if at the low level, you create a situation where you're playing with a completely different set of cards that nobody else is playing with, that enables outcomes that are very, very different. And so I, what I would encourage us all to try and do is try and understand enough about what is unique about Silicon One, why is it different? It is very easy to try and market around deficiencies in the industry, okay? I'm an engineer, I'm not a marketing person.
Maybe gives you a hint about the fact that we're not playing games and the fact that they would send an engineer to go and talk to people. The details matter, okay? I would just say that we are very far ahead of everybody else in the industry, and we're watching that play out over time, and it's quite exciting to be at, at Cisco as a consequence.
Well, great. Well, thanks for joining us here.