AI Data Center Networking Using Open Ethernet

0:04 so in a recent uh blog our CEO Rami

0:07 Rahim um really outlined why the AIML

0:12 workloads and their needs are

0:13 transforming high performance

0:16 networking the reason for that is that

0:18 if you look at the machine learning

0:19 workflows they fall into two categories

0:22 training and inferencing training

0:24 workloads are the most demanding because

0:26 you have thousands of gpus communicate

0:30 constantly constantly with each other

0:32 exchanging data so that they can train a

0:34 model and for that they need very high

0:37 performance High through food and almost

0:39 lossless operation to achieve that you

0:42 have to really crank up the networking

0:44 the second type of workload is

0:46 inferencing that requires very low

0:48 latency so you need to also focus on low

0:51 latency

0:56 networking the ethernet just like

0:58 Internet Protocol has been one of the

1:00 most open and successful standards

1:03 networking ethernet I worked on it for

1:05 many years has evolved to suit new

1:08 workloads new use cases again and again

1:12 and it has a very large ecosystem of

1:15 both vendors supplying products but also

1:17 expertise within the networking industry

1:20 now if you look at the needs of the AIML

1:23 training workloads of inferencing

1:25 workloads that I mentioned those can be

1:28 satisfied by the ethernet ethernet comes

1:30 with lots of bels and wh whistles or

1:32 features that can be used and tuned for

1:36 the to meet the needs of aaml

1:42 workloads and where Juniper comes in is

1:44 that we have been shipping products

1:46 based on open ethernet standard for

1:48 quite some time today we are shipping

1:49 our PTX spine switches and qfx switches

1:54 for these workloads along with what we

1:57 call fabric automation software collaps

1:59 ST which allows you to orchestrate the

2:02 end to end path from one GP to another

2:05 to achieve the high performance

2:09 n yeah so I think ethernet as I said is

2:12 a open standard for a long time and it

2:15 has constantly evolve and that's the

2:17 biggest strength the open ecosystem and

2:20 if you look at the standards bodies such

2:22 as I Metro ethanate and now there's a

2:25 new standards body be created called

2:27 Ultra ethanate Consortium just to

2:30 address the needs of high performance

2:32 and AIML networking needs these bodies

2:35 have made sure that ethernet continues

2:37 to stay open it has a robust ecosystem

2:40 that can meet the use cases and their

2:42 needs so I have every confidence to feel

2:45 that we continue to use ethernet as the

2:49 only open standard to meet the needs of

2:51 the

2:54 networking I think it's not enough to

2:57 have a high performance networking based

2:58 on ethernet uh you need to also the

3:02 orchestration of different network paths

3:04 congestion management flow control and

3:06 junipa networks has invested in a

3:08 completely new fabric automation

3:10 software called abstra that allows you

3:13 to orchestrate and automate the fabric

3:15 operation so what we do is we can not

3:18 only configure and set up the paths

3:21 necessary with necessary hooks uh to

3:23 take advantage of the ethernet features

3:26 but we also continue to monitor it

3:28 during its oper ations with Telemetry so

3:31 that we can build a Clos Loop automation

3:34 to make adjustments dynamically as the

3:37 network conditions or traffic conditions

3:39 evolve that's a big plus and very

3:41 important part of U high performance

3:46 networking so I think we are already

3:49 shipping 400 gig uh qfx PTX which is we

3:52 are on the CP of releasing our 800 gig

3:55 and we're going to continue to evolve to

3:56 the Next Generation to go from 800 gig

3:59 to 1.2 1.6 terabyte kind of speeds so

4:03 speeds and speeds and radius or radics

4:06 of the switches will continue to evolve

4:09 so that we can pack together in a small

4:13 cluster even a rack full of gpus lots of

4:16 high performance networking all based on

4:18 ethernet so it can be sourced from

4:20 multiple vendors using a robust

4:22 ecosystem that's where we are going

4:24 couple that with a robust automation

4:26 software you can really reduce the

4:28 operational cost and total cost of

4:30 operations for supporting such really

4:34 demanding

4:43 workloads