More Wide Area Networks

Intro: Welcoming AE Natarajan from Juniper Networks

0:05 [Music]

0:09 hey everyone welcome back to the 65 on

0:11 the road we are here today for an

0:14 exciting conversation talking about wide

0:16 area networks sustainability AI we are

0:19 diving in with a firsttime guest not

0:21 from a firsttime company we've got

0:24 Juniper Networks joining us we've got AE

0:28 AE welcome to the show this is your

0:30 first time excited to have you here uh

0:32 how you doing today oh wonderful uh it's

0:35 exciting to be here and exciting to

0:38 participate in this conversation you

0:39 know um it is it is it is all about you

0:43 know AI all of this excitement that is

0:46 going on in the network and the

0:47 inflection points it's it's really uh

0:50 it's really up there yeah I I really

0:53 believe you have a really important role

0:55 to play ae Juniper the business you're

0:58 in um we know that there's a lot of

1:01 excitement about silicon and compute

1:03 right now a lot of people tend to forget

1:05 that all this data has to be routed and

1:07 even we know when we spend a lot of time

1:09 talking about scale up infrastructure

1:11 and optical but a lot of stuff has to go

1:13 beyond these data centers we know you

1:15 know as it moves not just within data

1:17 centers but from data center to data

1:19 center from you know data center to

1:21 branch to edge to you know there's so

1:23 much going on and with all this there's

1:25 a lot of kind of these you know

1:27 tangental conversations you've got

1:29 energy and consumption and

1:30 sustainability you've got where does AI

1:33 not just come in in terms of all the

1:35 compute being used for AI but how does

1:36 AI actually help network more

1:38 efficiently so these are a lot of the

1:40 things that are going on really glad to

1:42 have you on the So give me uh give the

1:45 audience just the quick rundown you know

1:47 your EVP chief development officer what

1:49 does that remmit sort of have you doing

1:50 day in and day out at Juniper oh okay uh

1:53 it's uh it's it's an exciting job you

1:56 know uh I'll I'll I'll give you a quick

1:58 preview of what I did i was in the

1:59 networking field for a longest time i

2:01 thought okay networking is just building

2:03 more fatter pipes making things more

2:05 cheaper faster better uh but then uh I I

2:09 went on to build an application and the

2:11 application actually happened to be

2:12 video conferencing um uh the one that we

2:15 were using only to realize that the

2:17 network became such an important piece

2:20 of making this application win big right

2:23 if the network does not give the

2:25 guarantees of jitter throughput latency

2:28 delay any of those metrics in a proper

2:31 way we wouldn't be able to have this

2:33 virtual video conferencing and recording

2:36 uh be really good because it'll be

2:38 choppy voices uh choppy phases all sorts

2:42 of things brought me back to a moment

2:44 saying hey the innovation inside of

2:47 networks needs to be there uh brought me

2:49 to Juniper in a very exciting role to

2:52 build that innovation into our products

2:55 AI

2:56 native which essentially is important

2:59 for these days um uh not just AI with

3:03 compute and uh storage which actually

3:05 delivers training models and then

3:07 inference models and things that we talk

3:09 about how do you use it in every walk of

3:12 life the first part of it is how do you

3:14 take AI and AI models to help you with

3:18 what we build as network routers

3:20 switches and connectivity devices that

3:23 enable to provide the guarantees today's

3:26 network is using critical applications

3:30 if the network goes down you probably

3:32 have 911 calls not happening it is as

3:35 critical as that right so with that in

3:38 mind we really strive to innovate for

3:42 this modern era but with the advent of

3:45 AI that has come in it has actually

3:48 exponentially increase the changes and

3:51 demands that we have in the network

3:53 today so that is what is the excitement

3:56 that I'm here with Juniper because we

3:59 drive these innovations with our own

4:01 silicon and with systems and solutions

4:05 that really are future-looking and

4:09 enable to adapt to these changes much

4:12 more faster and much more easily so that

4:15 we can build it out that's great so

4:16 let's geek out a little bit and talk a

4:18 little bit about some of your your

4:20 products and as well as the chipsets

4:22 that you build i don't think a lot of

4:23 people fully attribute credit that you

4:25 guys are in the chip development space

4:27 you know custom chips um but you know

4:30 you've got ACX you've got MX you've got

4:32 PTX on the routers and then you've

4:33 you've built the Express 5 ASIC Trio 6

4:36 chipsets talk a little bit about how

4:39 they these these developments these

4:41 builds these products that you're you're

4:42 you're creating at Juniper are enabling

4:45 you know sustainable energy efficient

4:48 designs and and and bringing out the

4:50 performance required for these next

4:52 generation networks so so fundamentally

4:55 uh you know uh like you look at these

4:57 networks like we mentioned it right when

4:59 there was an internet boom connectivity

5:01 became very important right uh we

5:03 wouldn't have survived the pandemic

Juniper's silicon innovations for network performance

5:05 without uh without the network right we

5:07 were educating our kids we were ordering

5:09 our foods to entertainment to everything

5:11 with with the network and the changing

5:12 patterns right uh what drives us really

5:16 uh important is we start off with the

5:19 silicon silicon is a fundamental piece

5:21 of what we build here at Juniper so you

5:24 mentioned the Express and the Trio and I

5:26 want to tell you why we built two

5:28 different variations of the silicon trio

5:30 is a silicon that gives you flexible

5:32 pipeline which means it's very

5:34 programmable and it gives you

5:35 capabilities to adapt and adjust to any

5:38 new environments and applications that

5:41 people use typically Trio helps us power

5:45 what we call as the MX product line and

5:48 this has been one of the forerunners for

5:50 Juniper and the MX product line recently

5:53 is been the clear winner for being the

5:57 on-ramp to AI and AI clusters of large

6:02 major hyperscaler and cloud provider

6:05 when they want to onboard their

6:06 customers into their AI clusters these

6:09 are the devices that actually enable it

6:10 because it has the programmability it

6:12 has the capabilities the second silicon

6:14 that we build is the express silicon

6:16 which gives you the speeds and feeds

6:18 when we thought we did 100 gig that

6:20 would be enough it wasn't enough 200 gig

6:23 400 gig now 800 gig and tomorrow we're

6:25 going to talk about 1.60 pipes and these

6:28 pipes are getting filled faster and

6:30 faster with AI really coming into play

6:34 so the performance is really important

6:36 so Express drives that performance and

6:39 we build the PTX product line with that

6:41 performance and that scalability that we

6:43 do while doing both the Express and the

6:47 Trio silicon with our MX and the PDX

6:50 platforms we also take commercial

6:53 Broadcom silicon and build out ACX

6:56 products that actually are beneficial to

6:59 the customers because networks have

7:01 different needs so you need different uh

7:03 different devices in different places

7:05 that drives us to give a full portfolio

7:09 with AI native capabilities in the

7:11 portfolio so that is important for us

7:15 you touched upon another aspect of what

7:17 happens which is sustainability or power

7:20 reduction constantly we strive to reduce

7:23 power starting with our own silicon

7:25 silicon every time we improve upon it if

7:28 it is 2x the performance it is 2x the

7:30 performance with half the amount of

7:32 power that you need to use for it why is

7:34 that important the cost of power is

7:36 skyrocketing the cost of energy is most

7:40 important for us to actually have um

7:43 have a sustainable way by which we can

7:45 build and grow these networks so we

7:47 build it with our silicon we build it

7:48 with our systems we build it with our

7:50 software and we do phenomenal things

7:52 like green traffic engineering and

7:54 things like that to enable people reduce

7:57 power consumptions in networks even to

8:01 the extent of 70 plus percentage points

8:04 right so so let me layer this in you

8:07 also have done a very good job of of of

8:10 being you know among the leaders in 800

8:13 gig and then of course 800 GE tech how

8:15 does that sort of connect to everything

8:18 that you just talked about from a

8:20 hardware standpoint and then of course

8:22 with AI and intelligent traffic how do

8:24 you tie that all together to sort of set

8:27 a you know I I know the ambition within

8:29 Juniper i've spent time with Rammy i've

8:30 spent time with Manuj now I'm spending

8:32 time with you i know you want to you

8:33 want to be the standard in sustainable

8:36 high performance networking are you

8:37 seeing all this come together and get

8:39 you to where you want to go absolutely

8:42 absolutely uh it is uh uh the 800 gig

8:45 couldn't have been uh more perfectly

8:48 timed we lead and pave the way with

8:50 performance with our 800 gig products

8:53 we're the first to actually deliver 800

800 gig technology and AI cluster connectivity

8:55 gig routers uh that is becoming a lot

8:59 more important because if you really

9:02 take the AI clusters that people are

9:04 trying to build out um they need a

9:07 seamless way to connect these huge

9:10 clusters of GPUs when people were first

9:13 talking about 1k GPUs and 16k GPUs now

9:16 they're talking about 200k 250k and some

9:19 of the things are really million GPUs

9:21 right huge and and and guess what these

9:26 GPUs need to talk to each other to train

9:28 models or to inference things and when

9:30 you need to connect these GPUs together

9:33 you really need the network to be there

9:35 and the network has to be seamless and

9:37 provide complete reliability and

9:41 complete seamless no bottlenecks in

9:44 terms of messages and messages delivered

9:47 between the GPUs and the data that is

9:49 delivered so that they can orchestrate

9:51 and do really well so 800 gig has become

9:53 a powerhouse for people to actually

9:56 connect AI data centers AI data center

9:59 interconnects we also do a phenomenal

10:02 innovation on top of that in the van

10:04 where we use coherent optics where the

10:06 router itself drives these optics and

10:10 saves you power and energy and

10:13 complexity in the network removing

10:15 intermediary devices like repeaters and

10:17 others that you don't need anymore large

10:20 content providers are thinking of

10:22 spanning and the entire United States

10:26 with these kind of optics we call them

10:28 go ahead and CRZ plus optics which is

10:30 also part of our portfolio amazing to be

10:33 there at the right time and really

10:36 really enable this AI uh transformation

10:39 that's happening in the industry yeah

10:41 and we're going to talk about AI here a

10:42 little bit but it's I think really

10:43 important for everyone out in the

10:45 audience too to sort of understand that

10:46 you know you're really talking about

10:48 where copper and light meet right it's

10:50 that you know there's a there was sort

10:53 of this a bit of over rotation I think

10:55 in the market that everything was going

10:58 to go light and optical and you know we

11:00 know within the interconnects within the

11:01 clusters there's some extent of that in

11:03 in the rack but when you get outside the

11:06 rack and especially as you move you know

11:08 from you know maybe these biggest

11:10 hyperscalers you're seeing a lot even

11:12 within them you're seeing a lot of

11:13 standardization on Ethernet and um and

11:16 that I think bodess really well for what

11:18 you're doing but then of course you do

11:20 need optical in the right places for

11:21 certain types of connectivity and so the

11:24 fact that you're addressing all of it I

11:25 think some people really kind of

11:26 sometimes think juniper they think a

11:27 little bit more about traditional

11:28 networking but you're really attacking

11:30 all the vectors to address AI so let's

11:33 talk AI a little bit more um you're

11:36 using AI you're enabling AI from what

11:38 you just talked about you're also using

11:40 AI though um you know Mist uh Paragon um

11:46 and and you're doing this to drive more

11:47 efficiency out of the network And again

11:49 this brings sustainable uh outcomes it

11:53 lowers energy utilization you know talk

11:55 a little bit about kind of how you guys

11:56 are thinking about your AI uh

11:59 innovations um actually uh the way we

12:02 started this was uh Juniper was one of

12:05 the first to talk about what we call as

12:07 self-driving networks and self-driving

12:09 networks wouldn't have been possible

12:11 without using AI to enable people to

AI-driven networking for efficiency and sustainability

12:15 operationalize that network in a lot

12:17 more easy fashion right the first things

12:19 first about trying to do this is to get

12:21 visibility and observability in the

12:23 boxes that you're doing collecting a a

12:26 rich set of telemetry ability for us to

12:30 view that telemetry not just from one

12:32 device but from the network onwards

12:35 right and be able to actually utilize

12:37 that inference that build advanced

12:40 huristics build models to do this i'll

12:42 give you a simple example of what we did

12:44 about six seven years ago one of my key

12:46 engineers came to me and said traffic

12:48 engineering is actually the most complex

12:51 things that you can do in the network

12:52 that means the ability for you to

12:55 traffic load balance across um across

12:58 the entire network paths this is no

13:00 different than the freeways here in the

13:02 Bay Area right you use ways or something

13:04 it redirects you based on traffic

13:06 patterns and stuff like that the

13:07 complexity of that required us to build

13:10 AI models and to train them and we

13:13 trained those AI models what we

13:15 discovered was uh something interesting

13:18 the links were used only 44%

13:20 utilization and an average but there

13:23 were peaks and then there were valleys

13:26 the valleys made us think a little bit

13:28 more harder and say hey could I shut off

13:30 this link and save power could I shut

13:33 off this node should I should I

13:35 rebalance the routes should I rebalance

13:36 the traffic should I get better SLAs's

13:39 all of those things are how we started

13:42 with the genesis of how AI was built

13:44 into it we also have a recent innovation

13:46 where we've built the capabilities known

13:49 as an LLM connector that allows somebody

13:52 to go into our router and actually use

13:55 GI capabilities with their own LLM and

13:58 query their router and say "Hey you know

13:59 how do you want to do this?" And it was

14:02 funny because when we first displayed

14:04 this innovation in front of customers it

14:08 was an Italian customer and we started

14:09 typing in Italian the answers came out

14:11 in Italian people just loved it we did

14:13 the same thing in Japanese and now guess

14:16 what i don't need to actually have

14:20 translators or translations people can

14:22 actually train in their local languages

14:24 and people can understand and comprehend

14:26 the complexities of what we do in a very

14:29 simple way built into our AI native

14:32 products that enable this general AI

14:35 networking right and with that and with

14:38 the performance we also do the

14:40 networking for AI and there's a lot of

14:42 networking to be done a lot of routing

14:45 to be done AE um you know one of the the

14:48 things I think is very interesting that

14:50 Juniper's been working on is it's uh

14:52 universal routing platform um as

14:56 networking continues to evolve what we

14:58 route and how we route and where we

15:00 route you know needs more flexibility

15:03 you know we often uh you know we worked

15:05 within a lot of constraints but just

15:07 like how data needs to be more malleable

15:11 for AI you know you're seeing kind of

15:12 all these neo storage solutions that are

15:14 coming out because like we don't want

15:17 just file and block and you know we need

15:19 a basically a system that can read all

15:22 types of different data and access it

15:24 otherwise you know it take it just puts

15:25 more strain on everything network and

15:27 compute and you know so you're trying to

15:29 solve this for routing talk a little bit

15:31 about that yeah So uh uh this is this is

15:34 interesting because we we have been in

15:37 foremost in the innovation cycle for us

15:40 to be able to do um uh control plane and

15:43 admin plane and networks what we mean by

15:45 that is the ability for us to actually

15:48 drive uh how uh the network gets built

 

Universal routing platform for flexible networking

15:52 and how the routers talk to each other

15:54 to discover how the network is being

15:56 built and the links that are going on in

15:58 the network not only the links the

16:00 capacity of the links the throughput of

16:02 the links the current data that is

16:03 flowing through the links the ability to

16:05 provide feedback on those links in terms

16:07 of delay throughput jitter metrics and

16:09 other things that we can do we actively

16:12 put elements into it and we constantly

16:15 keep innovating to build on top of it

16:18 when we do this these innovations

16:20 actually are spread across every piece

16:23 of our devices that we build whether it

16:26 is the MX whether it is the PTX whether

16:28 it is the ACX whether it's a QFX or or

16:30 even the EX switches that we built all

16:33 of them carry this in a very seamless

16:36 way and they all react to it and build

16:40 total solutions that are needed for

16:42 today's network whether it is

16:44 enterprises whether it is large service

16:47 providers or cloud providers all of them

16:49 are able to use these elements to

16:51 actually drive the performance and one

16:54 of the embodiment of this is clearly our

16:57 MX304 the recent platform that we

16:59 launched has become a de facto for

17:02 people to be able to use this to onramp

17:05 a lot of um you know uh how you get

17:09 customers onto an AI um uh AI network or

17:14 an AI cluster which is amazing every

17:16 major hyperscaler uses this platform to

17:20 actually onboard their customers into

17:21 their AI clusters and AI native clusters

17:24 to be able to do both um uh uh

17:27 inferencing and other elements where

17:29 they can actually query do all of that

17:31 stuff that's amazing and this platform

17:34 is rocking and it's really really uh

17:37 catering to the needs of the people and

17:40 by the way this platform is built with

17:41 the trio which I talked to you about

17:43 that gives you the flexibility and and

17:47 ability to program and change it for any

17:49 growing and changing needs today we talk

17:52 about something tomorrow somebody comes

17:54 and says "Hey I need these filters I

17:56 need these changes I need these

17:57 different kinds of things put in the

17:59 packet." We can do that and it makes it

18:01 a lot more easier and simpler for us to

18:03 do yeah well like I said I think uh

18:05 fungeability is going to be very

18:06 important as network standards continue

18:09 to evolve as we sort of see ourselves

18:11 flowing back and forth between uh you

18:14 know high-speed copper and high

18:16 performing optics and of course we just

18:18 we've got so much change and it's

18:20 happening so fast i mean we're on annual

 

MX304 platform for AI network onboarding

18:23 cycles now one-year cycles which I don't

18:26 I I really don't envy the capex

18:29 management that these hyperscalers are

18:30 dealing with right now but I of course I

18:33 I am absolutely thrilled with how

18:35 exciting fast and how much money is

18:37 being spent to build out this AI future

18:39 ae um let's wrap up talking about

18:42 something positive something I think is

18:44 really encouraging you know I know that

18:48 during this sort of shift we've had a

18:50 pretty big shift in landscape political

18:52 shift and this is not a political show

18:54 but it's you know we we've seen a change

18:56 a little bit in in things like

18:58 sustainability and how much it's been in

19:00 focus um I think everyone most of the

19:03 tech industry agrees of its importance i

19:05 think they agree with its importance for

19:07 not just kind of the more altruistic but

19:09 also because you know if we want to keep

19:10 innovating we need to be thoughtful

19:12 about how we apply energy to solve

19:14 problems actually the tech industry's

19:16 always been very thoughtful about this

19:17 but as we build out AI AE we see

19:20 exponential growth that was like how do

19:22 we keep comm committed to this to this

19:24 zero emission like especially these

19:26 hyperscopes how do we stay committed

19:27 when all of a sudden we've we've

19:28 increased energy commit uh consumptions

19:30 by exponential amounts in like just a

19:32 year and the truth is they really can't

19:34 but you can you have basically doubled

19:38 down Juniper is saying no we're sticking

19:40 to it we're staying with our net zero

19:42 2040 uh commitment I'd love to hear a

19:45 little bit more from you about you know

19:47 how you've been able to to do this and

19:50 why you feel it's important important to

19:52 stay committed to this net zero by 2040

19:55 that you put out there so uh the reason

19:58 why this is important is because if you

20:00 really look at our uh look at how energy

20:04 costs have actually gone up and up and

20:06 up um whether uh and it is even more

20:10 significant in the rest of the world and

20:13 United States has started seeing it

20:15 right uh you see our um our energy bills

Juniper's commitment to net zero emissions

20:18 go up and if you go back and talk to

20:20 anybody who is building AI clusters uh

20:24 everybody thinks that those guys are

20:26 paying huge amounts of bills to uh this

20:29 small Silicon Valley company or a

20:30 Silicon company here in the valley but

20:33 they are also paying a much larger bill

20:37 a recurring bill yearoveryear to the

20:39 power

20:40 companies okay and if they if you

20:42 dissect that budget this this bill is

20:45 much huger our ability to tell our

20:49 customers and our people when they

20:51 deploy Juniper technology that they will

20:54 have the power efficiency to the finest

20:57 extent as they grow and scale up their

20:59 network and don't necessarily have to

21:01 burn more and more power is important

21:03 for us so that value and the commitment

21:07 to make that happen requires us to drive

21:11 drive it all the way from every aspect

21:14 that we do from our silicon which is the

21:16 most power

21:18 efficient tomorrow we're going to be

21:20 talking about space technology because

21:22 we have to launch these routers in space

21:24 or launch these routers in low orbit

21:27 devices and other things guess what

21:29 there is no power generators down there

21:31 you need to be extremely extremely power

21:33 efficient in space and guess our chips

21:37 are one of the most efficient ones that

21:39 is capable of being launched into space

21:42 that is the reason why we continuously

21:45 commit ourselves to building these kind

21:47 of chips building these kind of devices

21:50 solutions and continue to plow through

21:53 why we need to have power efficient

21:57 commitments that are important for us

21:59 all the way through if not today

22:02 tomorrow it is going to happen and

22:04 you're absolutely right everybody is

22:06 power hungry with AI clusters people are

22:08 consuming more power if we can everybody

22:10 pay attention to it in the longer term

22:12 it'll become a lot more easier for us to

22:15 build much more scalable networks and we

22:17 want to drive that innovation well I

22:19 want to thank you so much for taking

22:20 some time with me here on the 65 it was

22:23 great to sort of get the story and and

22:25 really tying some themes together i

22:27 think sometimes we want to look at

22:29 sustainability AI you know we want to

22:32 look at networking we kind of look at

22:33 these things as all these kind of silos

22:36 and I think the one thing we're learning

22:37 especially as we move from kind of this

22:39 first era of AI and ML to this

22:42 generative era to really this agentic

22:44 future is increasingly you know it's

22:47 it's outcome driven and the outcome

22:49 driven includes kind of the

22:50 incorporation of we got to manage the

22:52 power situation we've got to get data

22:54 from point A to point B um high fidelity

22:57 low latency you know we need enough

22:59 compute power to get this done we of

23:01 course need software that overarches all

23:03 this stuff and gives whether it's

23:04 flexibility in routing you know where

23:06 your engineers can design routing or

23:08 it's the ISVs being able to code on top

23:10 of this stuff and be able to get

23:11 solutions that work to people it sounds

23:13 like Juniper is doing a lot i know

23:14 there's a lot going on there you're in

23:16 the middle of a pretty big deal um

23:18 crossing my fingers you are able to get

23:19 that over the line but it was just great

23:21 to spend some time with you love to have

23:23 you back on the show um let's uh let's

23:26 talk again soon a Oh thank you thank you

23:28 very much really excited to be on the

23:30 show and I really enjoy it because like

23:32 you said it's all coming together and if

23:35 you don't see it as a full picture

23:37 you're missing the plot right so we're

23:39 we're glad we're seeing the full picture

23:41 and we're driving towards all that thank

23:43 you appreciate it very much and I

23:45 appreciate everyone out there for tuning

23:46 in to this episode of 65 on the road

23:49 great conversation here hit that

23:51 subscribe button join us for all of our

23:53 content we have a lot of great content

23:55 here on the 65 but for this episode it's

23:59 time to say goodbye i'll see you all

24:00 later

24:07 [Music]