Hybrid Is NOW – Why Ease and Speed Matter in Deploying the Right Infrastructure for AI

0:01 [Music]

0:04 [Applause]

0:08 hi I'm mansur Karam gvp of products for

0:11 data center here at Juniper Networks I

0:14 am fortunate to be able to talk to a lot

0:16 of you about data center needs and I

0:19 consistently hear the same

0:24 challenges everyone wants to move faster

0:27 to make their businesses more

0:29 competitive and the network is an

0:31 integral part of this you are all facing

0:35 huge operational complexities in the

0:37 data center there are fewer resources to

0:39 manage this complexity due to Chronic

0:42 skilles shortages you need to run at the

0:46 speed of the business yet any error

0:49 could lead to a disastrous outage

0:51 costing millions of dollars and there

0:54 are lingering supply chain issues and

0:57 other challenges in implementing hybrid

0:59 clouds

1:00 strategies ultimately what you need is a

1:03 way to build and operate private unram

1:07 data centers that are as easy to use as

1:11 the public

1:12 Cloud many companies are trying to

1:15 address these challenges with

1:17 do-it-yourself also called DIY

1:20 automation but this only seems to add

1:23 complexity increase technical debt and

1:27 increase dependence on scarce Talent

1:30 it's just not working for

1:33 them Juniper is in a perfect position to

1:37 help solving these pain points better

1:39 than anyone else in the industry whether

1:43 addressing the needs of data centers

1:45 with traditional application

1:46 requirements or building new data

1:49 centers specifically for AI workloads

1:52 Juniper has you covered this is thanks

1:56 to three sustainable areas of

1:59 differentiation that drive real and user

2:03 value first we simplify operations and

2:07 save time and money with an operations

2:10 first approach to design deployment and

2:14 troubleshooting abstra is the first

2:17 intent based networking solution for

2:20 data center operations and the only

2:23 solution that works across any vendor's

2:26 products this let you automate data

2:29 center fabric with maximum Assurance

2:32 greater efficiency and using fewer

2:36 resources second we provide open

2:40 flexibility that allows customers to

2:43 design their networks using proven

2:45 Technologies such as ethernet with Abra

2:49 you completely avoid vendor lockin for

2:52 better flexibility and cost

2:54 savings and third we have reliable and

2:59 secure TurnKey Solutions with endtoend

3:02 validated designs that include switches

3:06 routers automation software and

3:10 security this ensures confidence in the

3:12 choice of products expedites deployment

3:15 time and delivers on the promise of zero

3:19 trust Juniper's comprehensive portfolio

3:23 of qfx switches and PTX routers will

3:26 scale up to 800 gig and use a mix of

3:30 merchant silicon from broadcom as well

3:32 as custom silicon from juniper this

3:35 diversity provides even more flexibility

3:38 to you our customers and since taking on

3:42 the role we've committed to live deliver

3:44 the hardware platforms faster than our

3:48 competitors which is why we were the

3:50 first to announce the 800 gab ethernet

3:54 51.2 terabit per second Tomahawk 5 based

3:58 qfx platform

4:00 all of this results in up to 90% lower

4:03 Opex 85% faster deployment times and 9x

4:09 more

4:10 reliability and you have peace of mind

4:13 that Juniper is 100% interoperable with

4:16 leading gpus switch and data center

4:20 fabric fendors but you don't have to

4:23 just trust me when I tell you we are

4:25 driving real change and real outcomes we

4:28 have seen a one

4:30 143% growth in new data center logos in

4:33 the last 12 months when a company

4:37 deploys Juniper in the data center they

4:39 love it they're willing to talk about it

4:41 and there's no going

4:43 back I'm excited to tell you that things

4:46 are getting even better from here a

4:50 couple of months ago we made a huge

4:52 announcement to bring even more agility

4:55 Automation and Assurance to the data

4:58 center you just heard sudir talk about

5:01 all the great things we can do with

5:03 aiops in the campus and Branch

5:06 specifically with Marvis the industry's

5:09 first and only AI native virtual Network

5:12 assistant well we are bringing the same

5:16 VNA capabilities to the data center

5:20 Marvis VNA for the data center is the

5:23 first step on our data center aiops

5:26 Journey integrating the best intent

5:28 based networking with the best AI op

5:31 solution creates an Unstoppable

5:34 Force but talk is cheap let's show you a

5:39 demo in this demo you can see the

5:41 standard Marvis actions dashboard which

5:44 shows you the top issues and actions

5:47 associated with wired access wireless

5:50 access and

5:51 sd1 can you see what's new right there

5:54 in the middle is a data center leg for

5:57 the first time ever you get a complete

6:00 view of your endtoend network in a

6:03 single AI native dashboard if a problem

6:07 emerges like an application is running

6:09 slowly for instance you can easily

6:12 pinpoint which domain needs to be

6:15 addressed to fix the problem if you

6:19 double click on the data center leg you

6:21 see all the data center components

6:23 including switching devices connectivity

6:26 and so forth as you click further you

6:29 can see detail on every anomaly with

6:33 recommended actions for

6:35 troubleshooting all of this information

6:37 is coming directly from abstra so we're

6:40 leveraging the rich monitoring and

6:42 Telemetry that is already built into

6:44 abstra bringing it to the cloud and

6:47 presenting it in the Mist marvous user

6:51 interface if you need to Deep dive on a

6:55 data center issue you can click one

6:57 button to launch the abstra UI and

7:01 continue troubleshooting from there in

7:03 this demo we are also showing how a

7:06 simulated fault is detected wait a

7:08 minute that's not Juna CLI that's a

7:11 competitor CLI that's right folks this

7:14 is a multivendor network think about

7:17 this we have a multivendor data center

7:19 Network including competitor switches

7:22 that is being managed by abstra and

7:25 we're bringing all that full visibility

7:29 of that multivendor Network into the

7:31 endtoend view provided by

7:34 Marvis this type of endtoend visibility

7:37 delivers enormous value to our customers

7:40 and now we have a great tool to deliver

7:43 it even if there are no Juniper switches

7:47 in your data center well at least not

7:51 yet but that isn't the only iops

7:55 capability that we have added to the

7:56 data center Marvis VNA also has a

8:00 conversational interface that uses

8:02 generative AI for simple and seamless

8:06 knowledge based queries instead of

8:09 getting frustrated trying to sort

8:11 through technical documentation you can

8:13 just type a plain language question such

8:16 as what is a logical device and get

8:20 clear and concise answers and links

8:23 directly to the right document to learn

8:27 more these new aops capabilities are

8:30 very exciting but this is just the

8:32 beginning for Marvis VNA in the data

8:35 center over the next 12 to 18 months we

8:38 will be rolling out more AI native

8:42 capabilities to give you even better

8:44 automation agility and Assurance Juniper

8:49 is leading in both Ai and networking we

8:52 have followed a proven aiops blueprint

8:55 for the campus and branch which we are

8:57 now taking to the data Center our

9:01 competitors simply cannot touch

9:04 this but you cannot talk about Ai and

9:07 networking without also talking about AI

9:11 clusters just as customers need AI Ops

9:14 to simplify networking operations many

9:17 also need simple seamless and assured

9:20 networks to support AI workloads in the

9:23 data center we call this networking for

9:27 AI and it is a CR iCal element of

9:30 Juniper's AI native networking

9:34 platform I had the opportunity to talk

9:37 with Alan wle from the 650 group to talk

9:41 about this new market let's take a

9:46 look hey Alan thank you for joining me

9:49 today yeah thanks so much for having me

9:52 yeah know absolutely so this AI market

9:55 and especially when it comes to the

9:58 networking infrastructure piece uh how

10:01 do we Define it and uh also I'm

10:03 interested in its

10:05 size yeah so if we look at AI networking

10:09 there's kind of two networks there

10:10 there's a backend training Network some

10:12 people call this the AI Fabric or AI

10:15 networking piece uh the ethernet value

10:18 of that is about a billion dollars today

10:21 uh growing to 4 A5 billion in 2028 uh

10:25 there's also a front-end Network there

10:27 which is how we communicate to the rest

10:28 of the day data center that's a little

10:30 bit less than a billion dollars today uh

10:33 and growing to a 5 billion uh so we add

10:35 that up and we're talking about a$2

10:37 billion Market going to 10 billion and a

10:41 keger around 50% uh we've never actually

10:44 had a keger this high in data center

10:46 switching so remarkable growth and

10:48 driving the overall Market to record

10:50 highs every year wow yeah that indeed

10:53 that's a very very high growth rate and

10:56 um you you mentioned that ethernet was

10:57 one of the Technologies for the back end

11:00 um you know I believe you're referring

11:01 to infin band as the other alternative

11:04 perfect perhaps you can uh describe what

11:07 infiniband is and how they kind of

11:09 compare and

11:11 contrast yeah so infiniband is a

11:13 different network protocol out there um

11:16 there's pros and cons to both uh but in

11:19 general as we move forward ethernet will

11:21 become a larger and larger part of the

11:23 network uh out there I'd say infin ban

11:25 kind of comes from an HPC supercomputer

11:28 side of things and ethernet's obviously

11:30 the fabric of choice throughout the data

11:32 center and throughout most networks

11:34 right you know as they say never bet

11:36 against uh against ethernet and plus you

11:39 know it's a much larger ecosystem I

11:40 suspect also as there will be more GPU

11:42 options on the market you want the

11:44 technology that's kind of uh GPU

11:46 agnostic

11:48 right yeah absolutely as we get more

11:50 gpus and we're say more types of

11:52 training inference tier 2 training

11:55 ethernet does what it always does which

11:57 is a take over the market and becomes

11:59 the technology of choice out there yep

12:02 yep very true and so shifting gears now

12:05 just in terms of like the specific

12:07 characteristics of AI workloads

12:10 workloads and AI traffic you know again

12:13 specifically to networking um you know I

12:16 suspect there is a lot of fan in uh

12:18 elephant flows right given that you have

12:21 all of these iterations involved in AI

12:24 maybe tell us a bit you know what these

12:26 characteristics are for for AI workloads

12:30 yeah so you mentioned elephant flows

12:32 that's a big one we kind of have GPU to

12:34 GPU GPU to memory a connectivity GPU to

12:37 storage uh we then have a different

12:39 nodes transmitting all at once we've got

12:42 RDMA uh so this leads to a highly

12:44 latency sensitive networks and that

12:47 impacts your job completion time moner

12:49 I'm assuming you see some of this as

12:51 well when you talk to customer yeah know

12:53 absolutely and in fact you know it puts

12:55 a a big strain on the network right we

12:58 need like much larger capacity uh lower

13:01 latencies but also some capabilities

13:04 like you know you need kind of like on

13:05 the highway uh have an ability to

13:08 control congestion and uh and load

13:10 balance the traffic so that you don't

13:12 have the bottlenecks and it's it's a big

13:15 deal right for AI clusters if you don't

13:16 get it right these gpus are expensive

13:20 correct yeah absolutely like when you

13:22 don't get this right on a regular server

13:24 you're talking a few thousand dollar you

13:26 might be missing out when you start

13:28 talking about AI due to the cost of

13:29 those gpus this becomes hundreds of

13:31 thousands of dollars if your network

13:33 isn't tuned right and we're not even

13:35 talking about job completion time where

13:36 at it could delay your job completion

13:38 time in terms of training uh a data set

13:42 by what days weeks

13:44 potentially yeah on these new AI

13:47 clusters you're talking days weeks or

13:49 potentially months out there it's

13:51 something like you actually can't do AI

13:53 if the network isn't tuned correctly you

13:55 just would never get done training those

13:57 models so if you were to to to kind of

14:00 tell us the the kind of the keys to

14:02 success in this market if we want to get

14:04 into this market and be

14:06 successful what are kind of what

14:08 thoughts come to

14:09 mind yeah it's really about minimizing

14:11 the job completion time it's about some

14:14 of the things you asked in your question

14:16 about elephant flow latency load

14:18 balancing out there it's really about

14:20 creating a high performance high-end

14:22 Network something that's more similar to

14:24 HPC or what the cloud is used to than

14:27 kind of a traditional top Ira Network

14:30 where you weren't necessarily latency

14:31 sensitive yep indeed Alan thank you for

14:34 joining me today it was great to have

14:36 you yeah thanks so much for having me I

14:39 enjoyed it

14:40 absolutely jenniper is exceptionally

14:43 well positioned for this Market AIML

14:45 workloads in the data center are much

14:49 different than traditional workloads the

14:51 flows are large elephant flows they are

14:55 harder to load balance traffic is mostly

14:58 between between gpus and nodes tend to

15:01 transmit all at the same time and the

15:05 process is highly sensitive to packet

15:08 loss and Jitter these new flow

15:10 characteristics present complex

15:13 networking design problems suboptimal

15:16 Network tuning or configuration leads to

15:18 longer job completion times which which

15:21 could result in additional weeks if not

15:25 months for AI training not to mention

15:28 millions of doar of underutilized

15:31 gpus but rest assured Juniper is leading

15:35 with a combination of AIML features such

15:38 as load balancing and congestion control

15:41 and in operations first approach led by

15:44 abstra which fine tunes the network to

15:48 operate

15:51 optimally bottom line the result is an

15:54 optimal Network which is key to ensuring

15:57 that all the those expensive gpus work

16:01 together efficiently spend a little bit

16:05 more time and money designing the

16:07 network and save a lot of headaches on

16:10 the overall AI

16:12 application but don't take my word for

16:14 it I recently caught up with Udo wordss

16:18 Chief data officer at Fujitsu at the

16:22 world AI cans Festival he explained to

16:25 me how his building Juniper based AI

16:28 clusters to train self-driving software

16:32 take a

16:34 listen we started here with an executive

16:38 meeting in also com here November

16:41 last and we had 85 customers and

16:46 partners and at that time we had 12

16:50 pcc's in terms of Genera a we have an

16:53 own we have an own private GPT solution

16:56 where all the data remains in your

16:59 Center don't need a cloud service it's a

17:01 large language model you can feed in all

17:03 the information from your company and

17:06 you have information at your

17:08 fingertips now November till now is

17:13 let's say 2 and a half

17:15 months so Christmas in between and New

17:19 Year now we are facing 35 places so we

17:24 tripled the

17:25 numbers and end of this month I would

17:29 assume it's at least 45 to 50 it's

17:34 exploding first of all can we agree to

17:37 the topic saying a workload where it is

17:41 about self-driving class is one of the

17:44 hardest yes I think I agree yeah in

17:47 terms of um IO in terms of network Lo Y

17:52 in terms of GPU CPU whatever it okay I

17:56 can State out that I was involved in a

18:00 project where I can't give any names but

18:03 in terms of how to process these types

18:05 of data a real existing use case let's

18:08 call it like this um and we have

18:12 established an AI test drive which is an

18:15 infrastructure uh we have one in UK

18:17 London ATA and one at n in

18:20 Frankford um everything is um based on

18:25 true networks is based on zuza Rancher

18:30 Intel of course also inia elements as

18:34 well as net app storage so cool um and

18:38 it was just few weeks ago I was doing

18:41 the math and having counting the and

18:45 everything like this and we overachieved

18:49 expectations and this is not marketing I

18:53 have the slides in my laptop so we all

18:55 the utilization of iio of CPUs the gpus

19:00 memory even uh the the amount of

19:06 uh celius temperature of the car during

19:10 a specific process wow um so I can stay

19:14 out no that is not an issue with the

19:16 network but in this

19:18 respect our open solution with specific

19:21 features for optimizing AI over ethernet

19:25 avoid the challenges of infin band you

19:28 can Leverage all the benefits of a

19:30 proven technology including better

19:33 feature agility lower costs and no

19:36 vendor

19:37 lockin we have the best routing and

19:40 switching platforms for AI data centers

19:43 including the new qfx 5240 and PTX 102

19:49 launched this quarter and with TurnKey

19:53 validated Solutions you can get your AI

19:56 data center up and running faster and

19:59 with more confidence it will deliver the

20:01 results you need in fact we now have an

20:05 AI lab in Sunil where we are proving

20:09 everything out we are not stuck in

20:12 theory like other vendors we can show

20:14 you exactly how ethernet can match or

20:17 Beat infin Band performance for your

20:20 type of AI data center this is an open

20:24 invitation for all of you to come check

20:27 out our AI data center lab I am serious

20:31 I want to see you all in Sunny Veil

20:34 hopefully soon in short the jiper

20:37 approach will be critical to help AI

20:40 infrastructure move beyond the early

20:43 adopter stage today where it's mostly

20:45 hyperscalers building infrastructure to

20:49 mass Market where every Enterprise in

20:52 the world can have their own private AI

20:56 infrastructure working to solve their

20:59 particular digital transformation

21:03 issues wrapping up I think it's safe to

21:07 say that the quote unquote experts who

21:11 said everything will move to the public

21:13 Cloud got it wrong the future is hybrid

21:19 this means that not only are on-prem

21:24 data centers not going away AI is

21:27 driving even more demand for them in

21:31 fact my guess is that each and every one

21:34 of you right now are trying to figure

21:37 out how to use data centers to run new

21:40 Enterprise AI projects the data center

21:43 is one of the most exciting areas in all

21:46 of Tech right now if you're looking to

21:49 build a secure modernized data center to

21:52 reliably and simply scale Innovation and

21:56 embrace hybrid cloud and the AI

21:59 Revolution You are not

22:01 alone and from traditional applications

22:04 to AI clusters Juniper has your back

22:09 every step of the way our data center

22:12 Solutions are the easiest to manage most

22:15 flexible to design and quick to

22:21 deploy again don't take my word for it

22:24 real Juniper customers such as fast host

22:27 Yahoo and Advan are seeing an 85%

22:31 reduction in deployment time 90%

22:34 reduction in Opex and a 9x Improvement

22:38 in

22:39 reliability come to our lab sign up for

22:42 a demo and try a PC to see for yourself

22:46 I promise that you won't be disappointed

22:49 thanks for your

22:51 [Music]

22:55 time