7. The Evolution of Cisco’s Silicon Architecture with Jeetu Patel, Nick Kucharewski, and Mohammad Issa

Jeetu Patel, Nick, and Mohammad delve into our custom silicon strategy. While many see networking as a software-driven game, we discuss why many of the most significant breakthroughs are happening at the physical layer. We explore how Cisco’s unique approach to co-designing silicon alongside our software and systems portfolio is creating a level of performance and efficiency that simply cannot be achieved with off-the-shelf components.

Transcript

SPEAKER_02 0:00

Hello everyone and welcome to another episode of Inside Product. I have with me two very special guests today, Nick and Mohammed. The way I think about them is like Madonna, where they don't even need a last name. They're the Nick and Mohammed of the Silicon One team. And Nick and Mohammed actually run our um uh chip business, which is actually uh a very, very strategic part of Cisco. And Nick is on the product management side by background, Mohammed's on the engineering side. Maybe I'll have each one of you give you give the audience a little bit of a background. Um so Nick, we'll start with you.

SPEAKER_00 0:37

Sure. Yeah, so I'm an electrical engineer by training. Um I was at Stanford in the late 1990s, and all my friends were going off to internet startups. And uh I wanted to work in semiconductors, thinking that the internet was going to need a lot of chips. And I started out in '99 um at a at a company doing networking semiconductors.

SPEAKER_02 0:53

Did you did you always want to do semiconductors, or you actually happened to get into semiconductors at the time as a happenstance because that was your first job with some of the things?

SPEAKER_00 1:01

Um so that was that was the idea. Like I was focused on circuit design in college and then um started out my first couple years designing chips for networking, and then moved into product management about six years out of college. Yeah, but I've been working in networking semiconductors the entire time.

SPEAKER_02 1:16

And then you went to a bunch of different companies since then. Um and uh and the two of you work together as well.

SPEAKER_00 1:24

Uh that's right. We were at MMC Networks um beginning in 99, is when I started there, and Mohammed was on the team, and then we had a chance to work together at Broadcom a couple years after that.

SPEAKER_01 1:33

Awesome. And Mohammed? Well, uh I wish I was um decided to make uh decided to go into engineering in the 90s. It was in the 80s. So uh uh I graduated from Purdue University in electrical engineering and came to the Bay Area in uh the late 90s. I've been in the Bay Area since then in many startups, actually. I joined four or five startups. One of them was my own startup, and then I decided to join Cisco.

SPEAKER_02 2:05

And um, you know, we're gonna talk a lot about silicon today. And you um the three of us have spent an enormous amount of time together in the past two months um since uh I had the privilege of taking over this team. And thank you for all the patience you've shown me because I'm not a silicon guy by background, and you've been very, very good about actually getting me up to speed. But let's start with why is silicon so important um in networking and why is it so such a strategic differentiator for us to have silicon um as a core practice? And then we should also talk a little bit about onboard optics, which you own in addition to silicon. So start with you, Nick, and then we'll actually have Maham chime in as well.

SPEAKER_00 2:46

Absolutely. So, really at the end of the day, all of the systems that we're building have silicon inside at the core. And most networking companies, when they're looking at how do you um how do you build out that system, you have a choice. You can purchase chips from other companies, you can build your own chip as an ASIC, working with a third party as a partner to build that silicon, or you can do it yourself. And Cisco is is really in that third category, we're doing it ourselves, and Mohammed's leading the team that's that's making that possible. And the difference there is that by building it yourself, you can really innovate on the features at a much deeper level than if you're having to work with a third party for that. That means you can integrate the features with your software more closely, and you also have more control over the supply chain and that you're working directly with the founder, foundries, and supply partners. And um, and then finally, when you look at new technology innovation, it's about really having an advanced view and an advanced investment in those technologies that are going to make the next generation platforms possible. So Cisco is doing all of those things, and that's possible because we have that in-house team.

SPEAKER_02 3:48

And Mohammed, what would you add to that from uh from your lens on um you know the strategic importance of Silicon One with Cisco?

SPEAKER_01 3:55

Well, very few companies in the world uh can build silicon, especially big, large silicon uh chips that we're building. Very few. And it takes lots of Actually, namely four, I think, right?

SPEAKER_02 4:06

Four, four outside.

SPEAKER_01 4:08

Probably four, uh not many, as I said, you know, it's a handful. And uh it gives uh Cisco a strategic advantage by being one of the very few companies in the world that controls their own destiny all the way from concept to the supply chain to delivering a final product. This is not something that's easy or trivial. Many companies have tried to build these big devices. Everybody can design them. Everybody I can tell I tell people, everybody can tape out a chip. Yes. Taking it to production is a completely different story, and only very few can do that. Cisco is one of them.

SPEAKER_02 4:44

So let's talk a little bit about the range of silicon that we have and the kind of use cases that we solve and why this is important um for a different set of class of customers we have. So we have hyperscalers as customers, we have neo clouds as customers, we have sovereign clouds as customers, service providers, and then eventually all the learnings you get from those four really, really benefit the enterprise as well. And we've got um networking for both the data center as well as the workplace. And when I say workplace, I mean campus, branch, home office, what have you. Um what specifically walk us through the range and like what is the the defining physics marvel that is actually being achieved right now in the industry, uh, specifically as it pertains to what's happening with AI and the upsurge that you're seeing with AI.

SPEAKER_00 5:40

Sure, absolutely. So, so first of all, I think Cisco is somewhat unique in the breadth of the products that we're building all in-house. We have five product lines in full production and announced, and more in development on the way. And that spans from enterprise networks, both in the edge and in the enterprise core, enterprise data center, service provider networks, and then in cloud and AI, including scale out and scale across applications. So when you look at that, that's a range of almost 100 to 1 in terms of the fastest chip and the slowest chip across that range. But we're able to do that with a common architecture, which means we have a large amount of software leverage between those products, and then we have a lot of commonality when we're planning our products in terms of the technology that we can apply to build those chips. And um and so that matters because we're able to look across all of those networks, we can spot what's unique to those networks or what are the commonalities, and then we can build a long-term technology strategy with that vantage point across such a broad set of networks, which I think is is one thing that's very unique to Cisco. Um your second question, in terms of the uh, you know, the real technology opportunities that we see right now, or what's special in the semiconductor market now, is that AI is changing the requirements for both the hyperscale cloud as well as enterprise networks. Within the cloud, we see a significant increase in the requirements for bandwidth, lower latency, higher reliability, which is of course driven by AI training and inference.

SPEAKER_02 7:07

And then more efficient power consumption.

SPEAKER_00 7:09

Absolutely. So the really the question is how do you build a network comprised of thousands of GPUs or AI accelerators interconnected by that networking fabric that we built? We build that fabric. And how can you do that more efficiently? How can you do it more with better power efficiency? And then how can you enable your customers to take that to market faster? A lot of new opportunities for innovation that we're driving on. Over on the enterprise side, we also see a set of new innovations being driven by AI. Here it's somewhat different in that instead of building out the largest cluster possible, it's about AI happening on desktops, on laptops, agents running outside of the cloud. And it's really changing the way we think about enterprise networks. And that drives a lot of new requirements in terms of the software, which in turn drives new requirements for the silicon, which provides opportunities for new innovations in the product at the chip level.

SPEAKER_02 8:02

Would it be fair to say that even though the cycle time on these chips is compressing, and Mohammed, you're you're so close to this from the engineering standpoint where it used to be that a chip would take, you know, three, four, or five years to develop. Now you're starting to see cycle times close to 18 months on these chips. But what is happening is even though the cycle times are compressing, the lifetime value of the chip might actually increase and elongate over time. Because we have so many use cases that we can actually start to make sure that you can create variants of those chips that over time um that chip might actually have a much longer lifespan. So um do you agree with that firstly as a thesis and then talk a little bit more about uh well it it depends on uh which company is producing these chips.

SPEAKER_01 8:50

For Cisco, the answer is absolutely yes. Why? Cisco has a unique uh value that it brings to uh the end customer. We have um a programmable packet processing engine, and that by itself extends the life of the project, uh the the products. In the past, it you you know, everything not in the past, even now, uh companies build what we call hardwired packet processing, uh which means you have to know what the features are beforehand.

SPEAKER_02 9:23

Or you would have to do another tape out and make sure that Correct.

SPEAKER_01 9:26

So you if you know the features beforehand, you design for it, you verify for it, and you build the product. The feature changes, you have to retape out. In our case, we don't have that uh constraints. We have a programmable engine that allows us to add features even after we tape out and after the part has been in production for a while, a customer needs a new feature, we can add that feature. Now, of course, there are for everything there are limitations, but our engine is a very flexible engine.

SPEAKER_02 9:57

But would you agree with the thesis in general, Nick, that our chips that we have today, just like if you think about the H-100, it costs more today than it did two years ago. Um, the longevity of that chip is much longer, but the cycle time with which um GPUs are getting built is much, much faster. And I think the same is is true for networking silicon. So do you feel like we're gonna have, you know, instead of having a six to seven year lifespan of a chip, maybe there's gonna be a 10-year lifespan, maybe a 12-year lifespan of a chip. Is that a is that an accurate thesis to have?

SPEAKER_00 10:31

Yeah, I think so. Um one of the one of the um one of the elements that I think we'll see playing out over the next several years is that with such a um such large build-outs happening last year and this year, there's going to be a desire to extend the useful lifespan of that infrastructure. And so that means how do you actually maximize the ROI over an extended period of time? And that might mean deploying a mixture of equipment from this generation, maybe two years later, maybe two years after that, all forming a shared deployment.

SPEAKER_02 11:00

And they just degraduate the workload as they actually get less demanding for the kind of thing.

SPEAKER_00 11:04

Exactly, yeah, exactly. So the idea is not only designing for the latest and the current generation, but it's it's how do I actually look at a multi-generational approach to building out these deployments and making sure that you can get the most out of that equipment. Often that means interoperability and making sure the features across the network, when you look at the features tops down, there there's really a desire for consistency in terms of your telemetry, in terms of how you're actually doing the traffic handling. And the programmability is very special because that allows you to actually drive more consistency across different generations of the equipment, which in turn allows customers to extend the useful lifespan of the equipment that they're deploying.

SPEAKER_02 11:44

So we have made our own ASICs for a very long time. But 10 years ago, we decided that we were going to get into the full customer-owned tooling and tool set for chip design, right? Uh so direct to Foundry. Um walk people through what the benefit of Silicon One is. What is the benefit of having a single platform? And why is that good for an enterprise customer? Why is that good for a hyperscaler? What's the what's the end outcome that you get by having a single platform with Cisco compared to not?

SPEAKER_01 12:18

Okay. Um I'll take the uh the first part and um what Cisco used to do and what Cisco does now. Yeah. Uh we used to do uh we always did ASIC uh design or silicon development, but it was mainly uh ASIC flow. What is ASIC flow? ASIC flow means uh you do the front uh the architecture, the front-end design, the RTL coding, the verification, and then you hand it over to another company to build the silicon. Uh that uh comes with some uh price, some cost. Uh the cost is not just financial cost, it's also um it's not as optimized because the vendor, as an ex an ASIC vendor, wants to make sure they have enough margins for their manufacturing so they can keep increasing their gross margins and squeezing profit out of these devices. But Cisco ended up paying for that somehow because the the product is not completely optimized. You're leaving something on the table. Uh, with uh merchant silicon, that's another option, is you buy merchant silicon, but the merchant silicon has to have uh has to appeal to many companies, and you are buying features that you don't need. And still, again, the cost is also higher, and the power and other things because there are many features in there that you don't need, but you have to you have to take them. Yeah, with this COT customer-owned tooling, and the way Silicon One is done right now, is we control everything. We don't want a feature, we don't put it in. We replace that feature with something else that's gonna be. So the chip by definition is highly efficient because it's highly optimized for what Cisco really needs and the features that Cisco is driving to their customer, and you don't have that much wastage in the chip for uh uh for things that you don't need. Uh, in addition to that, obviously, it uh you have higher performance and in many cases lower power, and you control your own destiny in the supply chain. You don't have to depend on anyone except the foundry, and you manage your process shifting, the supply-demand wherever you need it.

unknown 14:34

Yeah.

SPEAKER_02 14:34

What would you add to that, Nick?

SPEAKER_00 14:36

Yeah, I think it's interesting. When when Cisco made the transition to COT and in-house silicon, um definitely it was driven by the benefits that you get from being able to control your own roadmap timing, being able to have a direct interaction with the foundry and the cost benefits and the optimization. What's really interesting is that that that decision ended up being pivotal to Cisco's participation in the AI transition that we're seeing right now. And so a lot of the success that we're enjoying in AI is really based on that decision to do our own our own chips. Specifically, what we see in cloud and AI, in AI in the hyperscale data center, is that the next generation of every AI build-out is really driven by what's possible with the next generation of silicon. So that the companies that have that technology in-house that are able to drive those new innovations are really working as partners with the hyperscalers to define that next generation platform. And because we're building the chips here in-house, we're looking at the technology we need, we can really work very closely to anticipate the next generation of AI build-outs. And that a couple years later translates into new business as we roll those out and we deploy those in production. And so that's really been a key part of our strategy.

SPEAKER_02 15:46

Well, what's fascinating is the actual amount of defying physics that happens in each one of these chip build-outs that goes on. Right? And I mean, talk a little bit about that.

SPEAKER_00 15:58

Yeah, so to start with, um, let's take the G300, which is a product we announced in February. This is our 100 terabit per second uh switch. And uh this device is a quarter of a trillion transistors, and there's multiple devices that are integrated that Mohammed and the team uh made possible um with uh with um with the design. And um and that really speaks to um an example of how the next generation of of the product is really driven by new innovations in core technology. When we look at what what it's gonna take to get beyond 100 terabits, it means new reinvention, a lot of new core technologies that come together.

SPEAKER_01 16:35

Yeah, this 100 terabit uh switch, for example, it's um You just said people have to kind of just take a moment and grok.

SPEAKER_02 16:42

100 terabits per second.

SPEAKER_01 16:45

Yeah. If uh about what um seven, eight years ago, twelve terabits was a big deal. Yeah. Was a big, big deal. We're at hundred terabits right now.

SPEAKER_02 16:57

And by the way, I don't think it's unreasonable to think we will eventually, in the not so distant future, get to one petabit of speed and about a trillion transistors.

SPEAKER_01 17:10

Like that is not like a that's not yeah, that's not an absurd ambition to have. No, that's within within reach. But that's within reach. Yeah, exactly. Uh we uh the technologies that we have to put together, the uh flows, it's not just tools and uh uh just architecture, it's every discipline is now extremely important and extremely specialized. From uh architecture to uh design uh and the verification effort is huge. Uh DFT is uh very complex.

SPEAKER_02 17:44

Explain for people that don't know what DFT is.

SPEAKER_01 17:46

Uh it's designed for testability. If you do not have good DFT in a device, you cannot produce it in high volume and good quality. It's extremely important piece in the past. Let's put it this way: in uh the early 90s, it was an option to add DFT into a device at that time. And when it was added, it was like, wow, it's amazing, and it was very very basic. DFT now is a very complex uh uh flow that we have to put in place, and without it you cannot mass produce and people don't give it enough attention. Cisco pays lots of attention to this.

SPEAKER_02 18:28

Why why is it more valuable for the customer to have a single platform Silicon 1 versus going out and making sure that every single chip they need is optimized for their particular workflow? Like what do they get out of it, Nick?

SPEAKER_00 18:46

Yeah, I think in particular we can look at the uh enterprise and campus markets. Um here you have a lot of different products of different bandwidths, different form factors. You might be talking about your branch office, you might be talking about the edge of the enterprise or the core or the enterprise data center. You really want to have products that are very well optimized for each use case. But when you look at the management of that network, you don't want a lot of different things. You don't want to aggregate nine different solutions that each have their own software programming model. So there's a lot of benefit from having a consistent architecture, a consistent programming model across all of those different form factors, but at the same time, having systems that are optimized for their use case in the network. And this is an area where we're seeing the in-house silicon really bring a lot of benefits for our enterprise and campus product lines. I think that benefit is going to continue to magnify and amplify with AI, because with AI, we see a lot of new applications inside of a campus network, inside the workplace network, because it's no longer being driven by the behavior of employees. In the past, we'd be looking at what hours are employees working. Are they bringing their own device and connecting to the Wi-Fi over the phone? When you start moving to agents in the enterprise, you're now no longer thinking about human behaviors, but what are the behaviors of those agents? That has a huge impact on the bandwidth of the network, on the security concerns. Um, and so the need to have the performance, the need to have that be optimized for different points of the network continues, but the ability to actually have full visibility across what's happening in your network becomes even more important. And because AI is moving so fast, it's really hard to pull together pieces from different places. So having that unified platform becomes critical. And that's really the answer to the question there. It's that the future is about velocity, but consistency in terms of programmability, in terms of managing that network.

SPEAKER_02 20:41

Two-part question for you folks is um one is describe for us the entire family of chips and what different parts of our estate we go out and cover, what kind of use cases we solve. And then Muhammad, geek out for us a little bit on why the way in which we've architected our chips is actually superior in the market. So let's start with the chips themselves.

SPEAKER_00 21:05

Sure. So um we use letters to to uh denote all of our different product families. And um, so I'll start from the highest bandwidth and work down towards the edge. So first we have the G product family, which is built for cloud and AI applications, and it's our highest bandwidth product line. This is the 100 terabit per second flagship that I spoke about. And you often often hear it referred to uh for scale out applications.

SPEAKER_02 21:29

Um the second problem scale out being that you can actually have multiple clusters within a data center that can get networked.

SPEAKER_00 21:35

Yep, absolutely. So the way to think about it is all data centers are um a network of GPUs or AI accelerators, and it's the networking equipment that makes those devices behave like a parallel supercomputer. And when we talk about scale out, it's about scaling out that that um those compute resources to form the cluster. And so that's really where the G. Series comes in handy, and that's where we're seeing a lot of success with multiple generations.

SPEAKER_02 22:03

Rows of racks. Yep. Just making sure that they've got a network.

SPEAKER_00 22:07

We picture rows of racks of thousands and thousands of devices that form that cluster is being connected by a device like the G. The second is our P series family, and this is targeted for scale across applications. So this is where you have a data center spread across multiple sites. And that's becoming increasingly important because with larger and larger clusters, you may actually exceed the capability of a single building to house that many accelerators. That could be because of power consumption, floor space, cooling, a number of different factors say that you want a cluster across multiple sites. When you go across multiple sites, you're talking about connections that span kilometers of distance. We're talking about coherent optics, like you probably spoke about on another session. And you also talk about routing capability, right? You're moving out of the switching domain and into long haul or carrier grade routing. And that's the product that we've built with the P-Series line. We also provide deep packet buffers, which are necessary when you're talking about those very long reaches.

SPEAKER_02 23:11

Because if a packet drops, otherwise, without deep packet buffers, you have to restart a training run, which could be extraordinarily expensive.

SPEAKER_00 23:18

Exactly. And that's really the key there is that um it's relying on the network to be to be reliable and to be to not drop packets. And when you're talking about running it across a very long-haul link, because it's going across this distance, you technically want to run it at a high level of bandwidth. You want to run it um hot, if you will, but you don't want to drop. And packet buffer is what allows you to do that, to simultaneously put a lot of traffic across it without dropping. Without dropping. And so with the P Series line, um our latest generation is the P200, 51 terabits per second, um, with those deep um uh in-package packet buffers. And there uh we're seeing a lot of success with that product line as well. Um next we'll talk scale across. That's for the scale across within uh AI data centers. Um next, we'll talk about um enterprise data center applications where we have the e family. And so this one has a lot of the attributes of the G family, but it's right size with the right features for that enterprise data center application. You would generally see slightly lower bandwidths, you might see some higher feature scale because you have um more of a heterogeneous type of deployment than the consistency you would see in the cloud. And it's it's really right sized and right featured for the enterprise. For the enterprise, for those applications.

SPEAKER_02 24:31

This is where our Nexus switches largely use, the e-Series family.

SPEAKER_00 24:34

Absolutely, that's where you would see that. And um and here there's a lot of features that we develop within the system level, within the software for the Nexus product family that are supported by that e-Series line. So this is where you see a lot of the advantages of Cisco having a common platform or a unified platform, uh, full stack with the software, the network operating system, software, the hardware design, and then finally the silicon. Like the e-Series really really plays a part there. Um next one we'll talk about is the A-Series. Um this is for enterprise uh access. Um and so you can picture um wiring closet switches, the the the uh the product line that really built Cisco's um uh product favorite.

SPEAKER_02 25:14

Campus brand switching for access.

SPEAKER_00 25:17

Absolutely. And so this is um this has been really the core of Cisco's enterprise product line for decades. And here with the A-line, um we have the optimization, but also the programmability that's needed to provide all of the feature sets for our customers, but also to provide that longevity in terms of new features over time. And then finally, we have the K-Series, which is our router product family, and this is targeted at service provider applications, but it's also used in the enterprise in some cases where you're spanning multiple sites, where you have to have um that routing capability um to interoperate with service provider networks.

SPEAKER_02 25:52

Yeah. And by the way, the P series is also a routing chip. Yep. As is the K-Series.

SPEAKER_00 25:56

Yep. And that's a perfect example where you have two devices with um a 10 to 3 in the bandwidth. Like, you know, when you're talking about these core optical router applications for between data centers versus enterprise routing, it's a 10 to 1 spread in the bandwidth requirement, but yet we have a common silicon architecture, which means a high degree of software reuse across those platforms.

SPEAKER_02 26:18

So, Mohammed, why is it that our chips are better than the competitions in your mind? What what what is special about these chips and what why are they so uniquely designed architecturally?

SPEAKER_01 26:30

We have uh built an amazing team. Uh we have attracted lots of talent from all over the place. Um we have uh flows that are uh that take silicon to production with you know the first uh tape out. We don't do two, three, four, five tape outs.

SPEAKER_02 26:49

Um why is that important for the customer?

SPEAKER_01 26:51

Because that's time to market. It's also based on the quality that Cisco uh is known for. We build um our devices with the highest quality, we don't cut any corners, we do all that's needed to get to production with full um full suite of tests that we do at every step of taking a chip to production. One advantage that our silicon, as uh Nick mentioned, uh it's we have common IPs in in the product that we use across the board. We do things like Certies and Certies, Max, even the packet processing, it's leveraged from the low end to the high end. So we have lots of leverage and reuse, which allows us to make sure that these IPs continue to improve and continue to be of high quality because they're used in many in multiple products. Uh, one advantage of the silicon that Cisco has that others do not have is we have, as I mentioned at the very beginning, we have a programmable engine. That's one piece of the equation that is very valuable for Cisco. We also have another very unique thing that no one else has: um shared buffer. A shared buffer architecture that we have in all of our chips, from all the way from the enterprise, the very low end, to the very high end. That shared buffer gives huge advantages to the customer. For packet from performance, as you said, packet drops, uh absorption of uh bursts, and so on. Very few companies uh can do what we do. So uh the combination of all of these aspects, the engineering, this the discipline in execution, the architecture, uh the IPs that we have developed, uh the quality that's enforced across all areas in Cisco, just gives Cisco a huge advantage.

SPEAKER_02 28:46

And I think one of the other big advantages that most people don't realize is we are a vertically integrated platform that is a co-designed full stack in the sense that because we make our own silicon, we'll talk about onboard optics in a second. We make our own systems hardware, we make our own systems software, we make our own platform for security, platform for observability, platform for data, um, our own models, our own applications, our own agents, and they all work in concert with one another, gives us a huge advantage that no one else has in the market. Um and so let's talk about onboard optics, because this is an area that, from an um IP standpoint, is pretty important. There's these things called co-packaged optics, there's near-packaged optics, both are on board. Explain why that's important and what we're doing over there.

SPEAKER_00 29:33

Yeah, it's a it's a great topic. And so when we look at optics, for the last many years, optics had typically been deployed as a pluggable module uh that would be inserted on the on the faceplate of the of the system. And there's a movement to bring that optics closer to the switching silicon that's making use of it. Umboard optics me or near package optics means placing those optical devices inside the system on the board, and co-package optics means bringing those optical engines right into the chip. So in this case, you can picture a device where there are no electrical traces coming out, there are actually fibers coming right out of the package. And the benefit of that is first and foremost power consumption. By having the optics leading leaving the device closer to the silicon, you can do that with lower power consumption. And that really matters when you're talking about hundreds of terabits of bandwidth. You know, being able to reduce the power that you're spending to drive those signals really matters. But um but this is probably a great one of the best examples of how Cisco's full platform approach, full stack approach, um, really makes a difference for product definition and execution. Um of the challenges with with co-package optics is not only building the device, but how do you build up the entire system using that device? The fibers leave the chip, you need to do fiber handling within the system, you need to do that in addition to the thermal mitigation like the heat sinks, how do the fibers work with the heat sink? Um, and then when you bring that all together, how are you going to manufacture that cost effectively? Then how do you do the reliability on that? If you're approaching that problem as a component manufacturer, think of all the complexity of working out all those engineering challenges and then manufacturing it at scale. Now, within Cisco, um, the team that's working on those type of problems are all daily on calls together, right? We're we're working side by side, we're we're really reaching out to identify different challenges and solve them really fast. And really it's about velocity in pulling these solutions together. And so one of the things that I enjoy so much working at Cisco is the fact that we have specialists in all those areas in-house.

SPEAKER_02 31:43

That's right.

SPEAKER_00 31:43

Silicon photonics developers, um, people with DSP experience in terms of the silicon. Digital signal processing, sorry. Digital signal processing, um, package design, um, the handling of the fiber, um, system level reliability, system level thermals, supply chain. And so we're able to talk about all of these topics very early in the development process.

SPEAKER_02 32:03

And it's not just the supply chain of the chip, it's the supply chain of the entire system.

SPEAKER_00 32:07

Of the entire system, of all those pieces that come together. And as we're seeing in in AI, um that's increasingly the critical factor. It's not only building the product, but how do you actually provide it at scale in very high volumes very rapidly? So by working together very early in the development phase, we can build a plan where the engineering is closely tied with the planning for supply chain and um in-field reliability.

SPEAKER_02 32:30

So um I could keep going with you too for a very, very long time, as we have done in many dinners that we've had and many meetings we've had that are day-long meetings. But uh if you were to tell a customer the things that are the most important things that you want to convey, what would those be from your perspective to the customer as it pertains to um our investments we're making in silicon and how that actually impacts systems, that eventually gives them better, more reliable networks that they can use so that they can enable themselves with the movement in AI?

SPEAKER_01 33:06

Well, um, I would say um I'm gonna whatever Nick just mentioned about the fact that when you get the silicon, the system, the software, the optics, all of it from one place, you know that it's been tested uh extensively and everyone worked together. It's not like it's co-designed. Everything is work everything is co-designed from day one. Yeah uh uh by by the team in one building or two buildings next to each other, there is one place that everything is coming together. It's very cohesive and it's very well understood compared to anything else outside where you have to take pieces from different places and figure out how to put it together. This is a key differentiator that we have. Cisco has um has been known for quality, and uh knowing that we are putting the optics, the silicon, the system all together, we guarantee the whole system together is of the highest quality. It's not about one component, it's about the whole system.

SPEAKER_02 34:11

Nick, anything to add?

SPEAKER_00 34:13

Yeah, so I would say that that Cisco at its core is really focused on building, deploying, and managing networks globally, worldwide. Whether it's a campus network service provider or cloud, it's focused not only on the hardware or the software or the management, but it's all of that. It's really how do you actually deploy and then manage these networks for the needs of the applications running on them. And for a customer looking at their next generation solution, it's important to look at that entire lifecycle of that network. And this is really where Cisco is focused. It's on having the right equipment with the right performance for today's needs, but also next generation. It's looking at how you're gonna actually manage that network for performance and security and visibility. Um, and it's also about planning for the lifecycle in terms of how are you going to mix this equipment with next year's equipment. Um, and then how are you going to um how are you going to manage your overall supply chain as you look at rolling out those networks and then um and then getting them to uh into market um on time. Uh this is this entire spectrum is where Cisco is focused. And we we look at that in terms of the design of our products, and we look at that in terms of how we work with our customers. And so really it's about um staying focused on that entire spike lifecycle of the network. And that's really the difference that we can bring, and that's what we focus on in the products that we're building and how we're working with the customers.

SPEAKER_02 35:43

I want to end with this. Um, take the focus away from Cisco for a second and focus on the market with what's happening with AI. What has surprised you the most?

SPEAKER_01 35:54

That Cisco is in the middle of this revolution in AI that will solve lots of things like disease and many other problems in in the world. Uh in addition to that, um, what surprises me how quickly now when we use AI at work, um we can use AI to help us debug problems, we can use AI to do some coding for us. I mean, it's just changing the way we do business on a daily basis. It's making us more efficient, which means that many, many more chips, hopefully. We will be able to do many, many more chips faster. And we need more engineers to keep driving these chips and getting them to uh to the finish line. So this is uh this is a surprise. I never thought that you know RTL used to be extremely important. Now RTL is important, but we also can get help from uh from AI to help us with it.

SPEAKER_02 36:49

But I will tell you, Mohammed, this one thing that I think a lot of times people misconstrue, and this is more true in silicon than anywhere else. The more you automate with AI, the more human bottlenecks will emerge. And so, in some ways, we will move the cycle time faster, but you're still gonna have a tremendous amount of need for human dexterity and for human intuition and judgment as you're building these chips. Because what I've learned about the architecture of these chips and how they get designed is you start with a blank slate and then you start to imagine the possibilities of what's gonna be the most efficient way that a packet can be processed on that chip. And um it's not something that, you know, it it requires a lot of instinct that actually is um it's gonna take a while for, you know, it's it's not something that AI is just gonna be able to, you know, uh do and replace humans. Like this is something that is gonna require a ton of engineers who are gonna get 100x more productive as a result of the use of AI.

SPEAKER_01 37:48

This is exactly what I wanted actually to stress that it removes the areas in our development process that are not uh they're not as innovative, for example. It's just you know how to do it well, fine, great. The machines can do it for us, and let's move you to work on more innovative solutions with the experience that you have.

SPEAKER_00 38:09

Yeah. So if I'm surprised by anything, I think it's the uh it's the magnitude to to to which the um to which these latest build-outs are happening. Um I think really this is really has been the dream. Like to go back to what I started out with is in the late 1990s, the internet is gonna need a lot of silicon. You know, as simple as that. And that when you first saw some of uh You're right about that. I mean that's and um but when you think about it, it in large part it took decades for all of this vision to really play out. Um and so what we're seeing is really a culmination of of what a lot of the the networking um industry has been planning for. And um a lot of the concepts that we see getting deployed that are allowing these big hyperscale data centers to get built up are concepts that we've been working on for for several years. And I'd say um if there's anything surprising, it's it's just that um it's it's really the um the magnitude in the scale. You know, seeing like tens of thousands or hundreds of thousands of connected accelerators um with an Ethernet fabric to connect them. Um it's it's really exciting to see that happening. But I think um the um the just the overall scale when you talk about industrial compute, I think that's not something I would have anticipated uh, you know, decades ago.

SPEAKER_02 39:25

You know what surprises me the most is the speed at which the market's moving. It's just fascinating to see how much defying of the laws of physics has actually succeeded. Where if you would have uh asked someone even 20 years ago, will we ever get to a petabit with a trillion transistors, they think you're nuts. And um uh now it's actually like, yeah, that's just a matter of time. You know?

SPEAKER_00 39:52

Yeah, I think you bring up a great point. I mean, for years we were talking about Moore's Law and this concept of just the exponential increase in the number of transistors in a device. But now with multi-dye, you know, two and after integration, dye stacking, now it's it's becoming a question of other physical um uh assembly challenges, or how do you actually build up these in in multiple dimensions? That's right. And so it continues, but it's pulling in a lot of cross-functional disciplines to make it happen. And I agree with you, it's moving amazingly fast. Um in and every two years, every three years is an opportunity for a whole new set of technologies to make the next leap possible. It's absolutely a lot of fun, but it's uh it's exhilarating for sure.

SPEAKER_02 40:33

Gentlemen, um I think Cisco is lucky to have both of you in these jobs, and um you folks are doing a fantastic um job at actually innovating for the future of AI. And so thank you for what you do every day.

SPEAKER_00 40:45

Thank you.