Maximizing Data Center Networks With the Power of AI

0:00 [Music]

0:05 hi everybody and welcome to our data

0:06 center video series uh my name is Dean

0:09 Sheffield and I'm joined by Scott sneden

0:11 um this is a a fireside chat Style video

0:14 series where we're going to try and

0:16 address some of the hottest topics uh

0:18 that we come across in our data center

0:21 experiences with our customers and

0:22 partners so we hope you find it

0:24 interesting um I think the first topic

0:26 uh we're going to address really has to

0:28 be Ai and ml mean it's it's wild it's

0:31 everywhere uh but what does it mean uh

0:34 for for data center operators people

0:35 building data centers people who are

0:37 putting their infrastructure in public

0:39 Cloud um there's there's a lot of a lot

0:41 of new buzzwords yeah a lot of stuff

0:43 going on so uh Scott what do you think

0:45 what's your first impressions well so at

0:47 Juniper anyway when we talk about AI

0:50 largely what we're talking about is mist

0:52 and and Marvis and the capabilities that

0:54 the team has built around making the

0:56 operational life easier making it easier

0:59 for Network people to run their networks

1:01 that's we we're leveraging AI really

1:03 heavily here to to lean into that to

1:05 deliver capabilities when we kind of

1:08 shift to Data Center and think about it

1:10 there's actually two topics so the first

1:12 is AI for operations and what we're

1:15 going to be doing with Marvis with

1:17 abstra with our intent-based networking

1:20 suite and bringing more AI capabilities

1:22 into that to just help operators do more

1:25 with less to do things faster to give

1:27 better insights and understanding so

1:29 that's a really

1:30 powerful tool that that a lot of that

1:33 exists today and a lot of that is

1:34 improving over time so so watch this

1:36 space the second topic around Ai and

1:40 data center is data center

1:42 infrastructure for AI workloads and that

1:44 is building very very large scale

1:48 high-speed switching capabilities and

1:51 fabrics to support these new AI

1:53 workloads um and that's also a really

1:55 exciting opportunity for us and there's

1:57 a lot going on there well let's break

1:59 that into to and let's let's have a a

2:01 chat about the the the AI Ops for the

2:04 data center operators I think some

2:06 people might be familiar with mist and

2:09 some of the challenges that mist helps

2:11 overcome specifically around

2:14 troubleshooting and finding out where

2:15 there's bottlenecks in the network in

2:17 the Wi-Fi network and in the actual you

2:19 know switching infrastructure yeah how

2:21 how does that sort of translate back to

2:23 the data center you think what what do

2:25 you think we can do there in the data

2:26 center I think it's what we're going to

2:28 see is is in machine capabilities to

2:32 make that troubleshooting more effective

2:34 we've already got an awful lot with

2:35 abster that's on the truck today a lot

2:37 of exciting things coming this year with

2:39 abster that make that even better with

2:41 visibility and monitoring and and uh um

2:45 tools to give you dashboards to really

2:47 understand and drill down to where

2:48 problems might exist and and try to

2:51 isolate and and troubleshoot things

2:53 before they become a major issue where

2:56 we take AI to the next level and and add

2:59 to that is to use some machine

3:01 capabilities to give you better Insight

3:03 better recommendations on where to go to

3:05 solve problems potentially even even

3:08 solving problems for you before they

3:10 happen I I think that's where we're

3:12 driving towards want to get to I I agree

3:14 I think

3:16 um the phrase meantime to innocence is

3:19 is a good one but also the faster you

3:21 can you know either isolate a problem

3:23 prove yourself innocent or at least help

3:26 remediate that problem yeah I I often

3:28 use the the phrasee the gift of time if

3:30 you can help people solve problems

3:33 faster that gives them more time to do

3:35 more you know productive things in the

3:36 data center what it's all about cool so

3:39 let's have a quick chat about the the

3:41 second thing you mentioned this this AI

3:43 infrastructure um I've done a little bit

3:46 of research and my my understanding is

3:49 that there's going to be uh obviously a

3:51 lot of data collection yeah uh we have

3:53 to gather all this data to then sort of

3:55 put into these training algorithms and

3:58 engines then at the back end of that

4:01 there has to be some other sort of

4:02 infrastructure that they call inference

4:05 infrastructure to try and you know bit

4:07 like chat GPT uh and B that that you

4:10 know once they've gathered the data and

4:12 trimmed the data down and then trained

4:14 it then we're going to see some new

4:16 things happening in the marketplace do

4:17 you want to have a a chat to me about

4:19 helping me understand the differences

4:21 between what that looks like from a

4:23 training of the data and then the

4:25 inference components to that yeah so the

4:28 um that sort of training of the data um

4:31 and and those those machine learning

4:33 clusters they're leveraging GPU type

4:36 CPUs they're um extremely high bandwidth

4:40 capable machines they operate in a

4:43 cluster so you've got multiple machines

4:44 communicating with each other those gpus

4:47 are extremely expensive um if you're

4:50 going to get your return on investment

4:52 from buying those machines and using

4:53 them the network that supports them has

4:55 to be non-blocking fully available um

5:00 very tightly managed for bandwidth and

5:02 capability so that it's always up and

5:05 these processes can run really quickly a

5:07 lot of that Market's in fiban today but

5:09 we do believe that it's going to move to

5:11 Ethernet there are a lot of reasons for

5:13 that around standards and multivendor

5:15 capabilities and we're investing heavily

5:18 in capturing that market so there's a

5:20 lot happening with with uh Juniper

5:22 around that and then as we look Beyond

5:26 just that piece of the cluster there's a

5:28 lot of of infrastructure to support

5:30 these workloads that looks like ethernet

5:32 that looks like a data center Fabric and

5:35 so what we have on the truck today

5:37 around high-speed switching um highly

5:40 available architectures automated abstra

5:43 driven intent-based Network

5:45 infrastructure to support any kind of

5:47 workload is very applicable to the

5:50 inference servers the sort of front-end

5:52 servers to these AIML machine you know

5:55 clusters so um yeah as far as building

5:58 infrastructure for this there's a lot of

6:01 stuff on the works to kind of get high

6:03 density high radics they call it

6:05 switching platforms um to support those

6:09 those backend needs and then on the

6:11 front end um what we have today is very

6:14 applicable and we're excited about it

6:16 excellent

6:17 so I'm thinking about use Cas use cases

6:20 now I I I did read something I think it

6:22 was like two weeks ago about some

6:25 Enterprises I think AT&T made an

6:27 announcement workday made an

6:28 announcement about

6:30 using AI type chat Bots and leveraging

6:34 training models

6:35 Etc I think there's some applicability

6:38 for you know some financial institutions

6:40 and some you know Healthcare

6:42 institutions where there's going to be

6:44 some sensitive data that's going to have

6:46 to be put into a a training model that

6:49 may be on Prem and not leveraging public

6:52 you know infrastructure that's out there

6:54 um any interesting use cases you can

6:56 think of around that well I I just like

6:59 we've seen with General Cloud workloads

7:01 there's a lot of sort of if not

7:03 repatriation at least reconsidering

7:06 where the best place is for a workload

7:08 and that carries forward with these AI

7:10 workloads as well um economically maybe

7:12 it makes sense to leverage a public

7:15 available learning cluster to get

7:17 started especially because they have

7:18 expensive they are very expensive but if

7:21 that's your business's crown jewels I

7:23 think we're going to see a lot of

7:24 investment within Enterprise to build

7:26 that stuff inh housee um so maybe Hybrid

7:29 models maybe there are you know leasing

7:33 for time on some almost bare metal as a

7:36 service type things but I I think we're

7:39 going to see a lot of Enterprises

7:40 building their own AIML capabilities in

7:43 I think it's very exciting in fact I've

7:44 seen some of the the forecast for the

7:46 addressable Market that that's going to

7:48 be out there and it uh it's people sting

7:52 billions of dollars of of of Investments

7:54 going into this space so extremely

7:57 exciting uh you know I think we're

7:59 extremely well positioned to partake in

8:01 this next boom and uh you know let's see

8:04 what we can do yep great excellent

8:06 thanks very much for for tuning in uh to

8:09 our first episode um we we'll be back

8:11 with another one very shortly appreciate

8:13 your time cheers thank

8:15 [Music]

8:21 you