Smarter Switching for AI with Cisco Enterprise Networking

Thank you, Raul. Um, my name's Kenny Lee. I am a technical market engineer, uh, from Cisco, obviously focused on a campus switching.

So if you look at the campus switching, the goal of a campus switching is provide connectivity to the different users, to the different endpoints. That goal has not changed. What has been changing or evolving are all the d all the endpoints as well as the applications running in the network.

I just kind of scan through right when I'm walking up here. A lot of us are using some sort of AI assistant tools, so H-I-G-B-T, germ knights and so on. If I look back into my office, right, we have a lot of people using, so for example, myself, I'm using CHE GBT to help me with my daily works.

We have developers using a visa studio with the copilot to help them do coatings, right? I also know people was using nano banana to help them, uh, do image creations as well as editings, right? So this is one of the thing that we are seeing very commonly in the ne um, in our today's networking, right?

We we're talking to a car manufacturings, right? So what is going on in the car manufacturing area? They were telling us that in order to automate this, they're gonna have a lot of high definition cameras monitoring the entire processing, and those data will be sending into some sort of AI processing and then take actions on those.

The second one we see is insights is gonna talk a lot more about wifis, right? We are seeing more and more customer adapting wifi sevens. The reason is, is a lot more bandwidth.

It has, um, use wider channels. That means in the same areas now you can cover a lot more user, much more denser users. So what is that to the networks, right?

So we have a lot more user, a lot more bandwidth that we need to take care of it from the switching side. Okay? And the last one, Raul has already mentioned that a couple more times in it, is quantum computing, right?

Typically, people look at this as like, why does quantum computing have anything to do with the networks? It is to do with the securities. Any cipher that you see is secure today might not be secure once quantum computing events.

So with all of those trend Cisco come out with, wow, I realize there is about five second delay. Next time I'm gonna press a button before I go through it. So with all of those trends, right?

What we did in the campus network looking at this is design. We are coming out with a new line of switches, right? This is the Cisco's Smart Switch series.

Within there, there is a 9,350, which is more focused on access layers, right? Whether you have a wifi seven or higher speed. So we have the 90 watts for any sort of, of a POE requirements, and we have the 10 gig speed on the copper to take care of any of the high speed devices on the uplink side, because with all this AI application, I don't know if anybody's look at it a little bit more detail, is the traffic become much more symmetric inside of asymmetric.

So previously you click on a button, I go to a web browser, I send very small amount of data's coming upstream, and I'm getting a large data coming back. Right now with the AI tools, you can have a lot more interactions with whatever you interacting with it. Now, the data become a bit more symmetric instead of always asymmetric.

Of course, the amount of data is getting a lot more. So the 10 gig going to the and host it address all of those requirements. Also, on the uplinks, we support a multiple of the hundred giga links, right?

What is that end result? In the campus network, you have a lower over subscription ratio. So you have a less chance of any congestions when you go talking to the outside.

Therefore, all of those AI application, which require performance and latency will be addresses in here. The second platform that we introduce is the 96 10. This is a huge beast, okay?

It has 10 slots, okay? The performance of a 25, about 25 terabytes today, okay? It just gave you an idea, is it can aggregate up to about 256 watts of a hundred gig today.

In the future, the chais is capable a lot more bandwidth. So in the future it can aggregate 512. That's what the cha is capable of.

I'm trying my timing. So I, I pressed a little longer, so now I can go to the next slide. The next one would Perfect a question on that.

Um, yes, you, you talk about that it being 250 600 gigabit, uh, links, can those actually be aggregated to be 400 or 800 gigabit links? So, um, the, the interface, the line cutter we are using in there, we have a four, a 400 gig purchase in there as well. So if you can use it for 400 gig, so within the example I gave it is the maximum of 256, right?

But out of the 256, if we, so today we only have a line cut that can provide two 400 gig per per slot. So the maximum we can get is a least 16, um, 16 of them. But in the future, if we come out with a newer 400 gig line cut, then think of it, every four port of the a hundred gig will be using as a 400 gig.

Okay? Thanks. Thank you for the question.

So at the heart of those, uh, driving this max switch is the Cisco silicone one as six. I'm sure a lot of you probably have heard that as well. The silicone one ASEC use a virtual procuring, um, infrastructures, right?

Um, why does it matter? I'll give you an example. Imagine in my floor, right?

We have a lot of software developers. They're using a copilot to develop the code, and then Friday at 5:00 PM is where they need to check in all the codes and say, everybody click a button and then all the codes we checking into the servers, right? Obviously there is a, an interface that that's going towards the server.

And since everybody download a large amount of data, so there might be congestion happening in the eagle side, but at the same time, hey, I'm done with my coding, but I still wanted to maybe do some documentation co write a couple slides on what I just did, right? And I'm using chey B to do this, but Chey BD is not going to the same direction where I'm going into the, um, server room, right? I'm going outside and talking to the chat GP agents.

So they are egressing out from a different interface. So even there is a congestion in one, our egress interface, but because I'm going to a different area, there's no congestion for my chat GPT traffic. So this is typically, um, when you talk about in the network design, this is what people are referring to is, you know, if you are blocking the other one that's called the head of line blocking, but in here we do not have that issues with the multiple virtual ques on any of one of the interfaces.

Also, with this architectures we'll be able to provide a much more granular controls for any type of traffic is going to the environment, right? Even within the AI traffic, you might have a different type of AI traffic. You might have an AI voice that I'm talking to it, right?

And you could, you might have just a normal inference and now go into some more details in, in those areas. So today, if you think of it in the campus networks, right? When we are designing network, there is a voice traffic, there's a video traffic, and you can prioritize them so that the voice traffic has that right treatments, right?

So what I'm talking to you, there's no latency, no jitters, right? You can hear clearly what I'm talking, but with the AI traffic, same thing within the AI traffic, each one of them might have a different requirements. Now with the um, uh, the s smart switch series, we have up to seven priority queues that you can use.

Now you can put all of different traffics into a different priority. So if I am just doing an upload, you can put it in very low priorities. But if I'm just using a voice to talk to an agent, I want a true a, like a live type, a live type of experience.

Like I'm talking to a live person, right? So you can put them into a higher priority queue and you will get the lower latency, the performance that you're looking for. So any, any questions?

Am I going a little too fast? No. Is, um, is this related to essentially, uh, data center bridging DCP?

Um, not quite. So in, in data center, the DCB, um, so I used to work in the data center Nexus. Um, this is more of a foundation of the queuing.

So we have a priority queues, right? Within the DCB, there is, um, a few other additional protocols in there to help make sure that when you, um, when you set up the, um, the pipe connection between multiple devices, there's always a, some bandwidth reserved for it. Right here we just really talking about within the single devices, we have multiple priority queue, and you could implement something similarly in there by using the right priority queue in each one of those devices.

Okay? Yeah, sure. So in the, um, so, So this, uh, this didn't actually tagging the, the content like DCP would be, It does not.

Okay. Yeah. So on each, so the, the priority in here is switch level.

Um, significant, it doesn't go between switches. Yeah. Okay.

The, what you're saying the DCP is between switches, it has that right end to end. Okay, thanks. Yeah.

Um, with the silicon wine, we are using two type of, uh, of buffer memory, the packet buffer memory. Obviously when a packet coming in, we have to kinda store somewhere, right? So we have some intelligent building as well.

So two type of packet buffer memory. One is what we call a sheer buffer, uh, sheer system memory, right? So this is very, uh, very fast, uh, shallow, not a, a big pool of the memories, right?

Obviously if I'm talking a voice or any of the AI agents, I wanted to have a very fast coming in and it quickly goes up. But if I am just doing some sort of, uh, inference, right? Um, uploading a large amount of data coming in, it could take a little longer, that's fine, but I don't want the packet to be dropped.

Even if in my previous example, everybody's checking into the code and into my server, it's a little take a little longer. It's okay, I don't wanna be dropped. And then this TCP session get restarted and I have to restart the session again, right?

So there's an intelligent building here is for all of those latency sensor traffic, it just use the shared buffer memory. But any of the larger packer sites, which is normal bulk data, right? It use what we call an XVM, high bandwidth memory, those could go to take a little longer.

So at the end of the day, right, you have the performance you're looking for, okay? Also, with all of those, um, new application, going to the data, uh, data going on the campus, right? You have the users, you have the different application.

There's also a lot of talk about this, a AI agent talking to another AI agents, which means all of, suddenly if you're looking at the campus environment, I have a lot more endpoints, a lot more users talking to each other. So you wanted to provide the visibility as well as some security policies on them, right? To do this, right?

We are using something called hcam hash cam. It is a combination of tcam plus SRAs, tcam. A lot of you, anytime you do the networking, you hear about that terminology a lot, right?

It's very fast, but it take a lot of resource to implement it, okay? Which means it's gonna be cost and also it's gonna be generate a lot of heat. Sri on, on the other hand, it just a very direct match on it, right?

So it use a lot less resource, less cost and less power. So the combination of the two, we can use it, alright? Make use of each one of the benefits and it and get to the result that you're looking for, okay?

With the, um, with the H cam, we, today we implement two type of features in here. One is net flows, right? NetFlow is obviously providing the, the, uh, visibilities, right?

Um, NetFlow simply just what host talking to what host or whatever the thing you're looking for, right? So we can implement those in the hat in the SRI pretty easily. The second feature that once you have those vis visibilities, right?

You wanted to implement maybe some policies who can talk to who, which AI agents should not be talking to, which AI agent or which user should not be talking to, uh, which endpoints and so on. So this is where ACL comes in. You can implement those in the hardware on the platform itself, okay?

I give you an example, right? If I have a thousand line of ACL previously, that means I need a dozen inches in the tcam, which is gonna use a lot of hardware and resources. Now that match of what hose, what IP it is, I can implement those into the sram and then only the action pack I can use the tcam.

So you kind of need, you know, going from a thousand needs for the tcam, I'm down to about three, maybe three inches in the tcam. So dramatically reduce your power and cause in this environment, um, so Raul will talk a lot about this, but on the, um, this is about the, um, the security, right? The security is fuel in the network.

This is not only on the routing platform, also on the switching platform. So the very fundamental piece of this security foundation is a trust anchor modules, right? This is a piece of hardware that we put into every single one of the Cisco device.

What, what do we have that is in there? It has certain cipher, it has certain signatures is in there. So when the system is, is putting up, right?

Not talking about just the OS, right? Even in a very lower level and the round one, we check those cycles and make sure that, okay, this is a genuine Cisco product. And when you're booting up any of the round one image, is this a genuine, anybody have modified that, right?

Those will be, will be checked as in there and not if you move up to the layer now between the, the devices, right? You have the transport layer. So this is where the IPSec max that comes in.

Now, if you look at the network as a whole, right? As somebody who is asking questions about zero trust, um, networks is in here, right? How do I make sure any of the endpoint connected to the network or any user connecting the network, right?

We don't trust them at the first, right? You have to go through authentication. And once they've been authentic, uh, authenticated, right?

What are the policy, what are the things that you implement in there? We use a security group chat to identify each one of those. Like, hey, this is Kenny, he's belonged to, um, marketings, right?

And it would be someone else who belong to a finance, right? So each one of those different group, and we use that to implement policies, um, in conjunction with the, um, ice engines, right? The ident ident identity service engines.

We, uh, those, the next couple slides I'm gonna go through pretty quickly is, right, so because of the confirmed computings, right? How do we enhance those cryptography that you using in the existing, um, whether it is Secure boot or whether it is IPSec, um, whether it is Max Act, right? We need to enhance those protocols, but the foundation is you need to be hardware ready.

The hardware need to have those, um, capability built in, right? And subsequently for the IPSec, for the Maxx, you can just, uh, deliver that capability with an additional, um, software releases and so on. I talk a little bit about the, the security group text, right?

This is how you deliver the end-to-end security within this environment. So Cisco do offer a secure fabric solution. This is a turnkey solution that you can use with a, um, side talk about it with the catalyst centers and so on, right?

Within there it's integrated with the, uh, identity service engines, right? Any device, any user comes in, it will go through the profiling, right? Once it go to profiling, we know what good taxes belong to, right?

And this process is continuous, right? There is a learning and the ice engine will get a continuous update, right? Let's say there is a new device means probably gonna talk about a different endpoint connecting to the wifis and so on, right?

When there a new one coming in, right? We will learn what type of device they are and what group that they're supposed to go into, right? So this process is continuous and the AI in the back that is keep learning those, okay?

With the, on the, um, the nine k, uh, s Smart switch families we have enhanced the CPU and the memories. Okay? So now you can have host a workloads closer to the edge, right?

Some question was asking about, Hey, what happened if, what about the MCP servers and some of the AI agents right? Now you can hold some of those AI agents much more towards the edge, right? Any traffic that's coming in, you can initially, you can process this much faster.

And also because this is more distributed, right? You can still have your more central AI agencies in here, but they don't need to pull into each one of them, right? I think Shine can also kind of touch on it is I have a North Bank controllers in here, right?

Do I need to ping each one of the endpoints or I can just talk to a, a bit more aggregated right? Pharma switch side, I'm aggregating those information. You can just talk to these agencies in here, right?

The capabilities there. Um, so today some of the, um, application that you can host us on there is like 1,009. Should I use that as an example that you can use to troubleshooting it.

You can host a thousand nines in there. Space means safe far, we can talk about right space gave you a location of a different endpoints with that environment and you can get that information as well. So all of those is capable with these smart switches.

And we also enhance the amount of data that you can use for those application hostings, right? Used to be, you know, after a few gigs on the 9,300, now with the 93, 93 50, it's up to, um, 20 gigs, uh, with two, you know, two times 10 gigs basically. And question about the CPU choice here.

It's been a four core for a while. Is there any potential for it becoming an eight core or higher on that CPU? Okay, um, on the access platform.

So we are looking at, so within the 93 50 families, right? There could potentially be other models that will have higher ups, right? Okay.

Yeah. Um, now you design it a network that is secure, right? High performance and so on, which you need to operate the networks, right?

And operating a network involved in software updates as well, okay? Because it could be a security issue, it could be, um, just a feature that I needed, right? Um, 9,300 we do support XFSU, right?

Is for any of you who have familiar with it is we first introduce it in 17 point threes or so. Initially, uh, we drop the, uh, outage time to 32nd and typically we drop it to, uh, five seconds, right? With the most recent release on the 93 fifties, right?

We're doing further enhancements is in here. Okay? So XFS U the way that it works is I separate the control plane in the data plan, right?

The data plane will continue to forward traffic while I'm upgrading the control plane, right? I could take five minutes or so, and the control plane doesn't matter. Once the control plane is done, right?

I talk to all my neighbors, I can get the latest information now, then I can reset the, um, the data plane. The data plane. In this case it just the ASEC itself, right?

So previously we just Reset. That's what the B stands for. That that's like you, you have BPDU that that's the plane you're talking about.

Yeah. So sorry, I'm gonna move the control plane and the data plane. Okay?

So I, I'm just looking at the third bubble where it mentions BPDU. BPDU. Okay?

So I'm gonna go into that in just a second. Okay? So we separate the two, uh, two planes, right?

So when we reset the data data plane, you have a some amount of, um, traffic interruption. And this is why I said initially we take it from 30 seconds and now down to five seconds, okay? But with the 93 50 we further enhance this, right?

Even with in the control plane when I need to, um, restart the control planes, right? It used to take a lot longer even though there's no traffic interruption. But now we can just reload the kernel and that take it from three minute down to one minute.

Okay? The second one we did in here is when we resetting the forwarding plane, we don't just reset everything in the forwarding plane, we only reset whatever the features enables you there. So that could further nail down to your dis um, traffic disruption type.

And the last bucket, uh, BPDU, um, is specifically, um, is specifically, um, the BPDU, right? Um, spending tree within the as itself that we do have a small processing power system there, okay? When the control plane goes down, we all flow this to the NPU that within this asic to tell my neighbor and say, Hey, I'm still here.

And you know, the, the BBDU processing, right? The hollow stuff, and I'm still talking from my neighbor routing protocol, has a grateful restart. So that's fine.

I tell my neighbor, Hey, I'm gonna go down for a while, I'm good. But spanning tree doesn't have this capability. So what we can do in here is we offload this to our CP our, our little npu within the, uh, a sickle cell so that my neighbor will always think that I'm still here, right?

So the end result is we're gonna have a much shorter, um, outage time. Yeah. So I talk a lot about what are we doing for ai, what about the AI on the networking, right?

Should I talk about, hey, we have the AI agents, right? And we have the AI canvas, or sorry, the AI assistance and the AI canvas with the AI assistance, right? What we can do today is like if you use some, if any one of you're familiar with the Meraki dust sport, you can go today, right?

You go into the assistant, it has many tools to help you with your day, zero day one day end operations, right? Day zero typically you do the onboarding. Mera does those pretty easy.

But if I wanna do some sort of inventory and looking at what platform is currently there, hey, any platform that I need to upgrade, hey, maybe I should go to the 1950. I can get some information on the 1950 right? For troubleshooting, right?

A simple example, we were on spinning trees, so I can quickly find out which one is my route when the spinning trees, right? Has the route been changed, uh, during the last 24 hour, right? This will help you troubleshooting.

Hey, my network's all stable, it's all good. No, my network has some issue that I need to take care of it, right? For the, um, a day end operations.

Right? Hey, any, any one of those device have a new image I should upgrade to, right? Um, maybe for compliance, all of those, you can be get it done with the AI system in a much shorter times, right?

And you can type something in and just wait a little bit, get the cup of coffee, and then come back and get all those results. So the key takeaway, it's, um, so we introduced the new K nine K switch, high performance, ready for the, um, the AI arrow in the, in the AI campus environment. And the security is everything is infused in the device itself, right?

And we are quantum ready for those environments. Any questions? Yeah.

Uh, does, um, do these switches have any capabilities for any of the new ultra ethernet features? Uh, particularly both, um, multipathing and out of order delivery as well as you were talking about tele, uh, the telemetry on them. Are they taking into account any of the new telemetry features that are being proposed in ultra ethernet?

Or are they software capable of being upgraded to do so? Uh, sorry, the first question you have, it Was, uh, it's, are you, are these switches capable of, uh, being upgraded to work with ultra ethernet? With ultra ethernet?

Yes. Yes. Ultra ethernet.

Yeah. Yes. It's a, you know where it is?

No. Okay. Actually, okay.

And then question that is it, it's okay. Web ethernet, I heard the TOM one. Ethernet.

Ethernet, okay, I'll, yeah, exactly. Yeah, yeah. So I, I'll just say that the switches that we are talking about is for the enterprise networks specifically.

So meaning, you know, campus networks where you have the wifi connected and the endpoints connected. Um, so we don't have, um, you know, so we, we don't play in that space. I think the, the key message that we are trying to drive here is as AI comes into the network, obviously data center has been the focus and there is a ton of AI processing happening in the data center.

But think about the rest of the enterprise networks and, you know, hotel networks, campus networks, the traffic patterns are changing. And for that traffic patterns, like you mentioned earlier, we, the, the network has to adapt to that. And for that, we have, in the smart switches, we have new capabilities that are actually, um, built for AI traffic.

And that's, that's I think from AI perspective, that's, that's a key point that we wanted to drive. Um, in addition, you know, the world is evolving very fast with quantum computing and other security threats. And so you need, uh, in infrastructure and networking infrastructure where, where, uh, that is ready for these new challenges in the, with the, with the new AI traffic.

Think about the new AI traffic that's coming into the network. It's a lot of traffic and it's a new type of traffic. My, my daughter keeps talking, uh, to the HR GPT.

That's a very new type of traffic that didn't used to be before. Right? And so that's, that those are the switch and infrastructure that we are talking about here.

And I, I do have a question about, a little bit back about the memory allocation between Yeah. The, the hash cam tcam, Andra. Um, is that a dynamic allocation or is that still kind of from a, from a former term, a fixed allocation where I've applied the memory maps on the switch and it's permanent until I change it, and then I have to go through that reload cycle.

Ah, You're talking about the, um, like SDM template type of Thing? Yeah, the templates. Yeah.

Yeah, Exactly. So we, today is fixed. Okay.

Right. We allocate half of the memory for the, um, uh, net flows and the other half for the ACL. Mm-hmm.

Our goal is to be able to allow you to customize it. Okay. For example, if I need a lot of, of, of one way, then I can customize it.

Okay. Future customization. Future customization.

Perfect. Yeah.