# Silicon Photonics Switching: From Technology to Deployment

https://www.youtube.com/watch?v=gLXt_8e1cEU

[00:01] Right, thank you, Richard, for the introduction and for the invitation and for the opportunity to be here and speak about the silicon photonics and how can we use this technology to to introduce optical switching at massive scale for next generation AI data centers.
[00:18] centers.
[00:20] It might not be necessary given the audience, but if we go super fast on the typical architecture on the data center, what we are starting to see is a real network challenge.
[00:29] We have a a single compute unit.
[00:31] In the past, it was a single GPU or XPU.
[00:34] These servers have been growing and growing with 8 GPUs, 72 GPUs, and even plans to get that into hundreds of GPUs per cabinet.
[00:44] These cabinets are typically wired up in the electronic domain with copper in order to increase the computational power.
[00:52] Once you have reached to the limit of a scale up, then you put another cabinet and another and another until you have enough computing power to develop your
[01:01] power to develop your your computing needs.
[01:03] your computing needs.
[01:05] These extra cabinets are typically wired up in the in the electronic domain.
[01:07] up in the in the electronic domain.
[01:09] Actually, it's an optoelectronic domain already.
[01:11] You start using optical interconnects that requires the movement of signals from the electrical domain to the optical domain.
[01:14] of signals from the electrical domain to the optical domain.
[01:18] You go to the own optical fiber and then you can start creating this type of pods.
[01:23] creating this type of pods.
[01:25] This is typically known as the scale out domain with approximately one order of magnitude lower bandwidth and higher latencies, but the goal of the industry is always to increase and and increase the computational power first on the scale out domain.
[01:38] [snorts]
[01:40] If we take a look to the conventional architecture of data centers, there's not such a conventional architecture.
[01:45] not such a conventional architecture.
[01:47] There There are actually different ways to get there.
[01:49] A typical one is this spine leaf architecture that aggregates the computing pods uh relatively complex uh network.
[01:57] uh relatively complex uh network.
[02:00] One thing that we have been seeing in
[02:02] One thing that we have been seeing in the industry is kind of a pivotal change.
[02:05] The industry is kind of a pivotal change where some hyperscalers started to propose the transition from these um electrical uh fully electrical switches to explore the option of putting optical switches in the network.
[02:18] In the upper part, you see the pioneer case of uh hyperscalers like Google who which bet that a long time ago, several years ago, on the option of having uh of replacing actually the upper layer of switches uh typically done with high capacity electronic switches by um optical switches.
[02:41] This allows you to remove one of the hopes uh from optical to electrical, electrical to optical domain, the substitution of hundreds of millions of transceivers, and more importantly, there are several applications that uh the industry has been exploring so far uh in the spine layer replacement case, in case of Google as well with the TPU enabling their
[03:05] enabling their um um highest performance uh TPU systems.
[03:09] highest performance uh TPU systems.
[03:11] In the bottom side, you see an another interesting case which is
[03:13] interesting case which is I mentioned before, the the key goal is
[03:15] I mentioned before, the the key goal is first, you increase the scale up domain,
[03:17] first, you increase the scale up domain, and once you are done with the scale up
[03:18] and once you are done with the scale up domain, you you you get the scale out
[03:20] domain, you you you get the scale out domain to to continue doing additional
[03:23] domain to to continue doing additional parallel processing, especially data
[03:24] parallel processing, especially data processing.
[03:26] But tensor uh level uh parallelization, pipeline
[03:28] uh level uh parallelization, pipeline parallelization most of the time require
[03:30] parallelization most of the time require having a a single domain, high scale up
[03:33] having a a single domain, high scale up domain.
[03:35] domain.
[03:37] The exploration of of optical switches, not as a direct substitution of uh the
[03:39] not as a direct substitution of uh the scale up switches that we have today in
[03:41] scale up switches that we have today in in the in the electronic domain,
[03:43] the in the in the electronic domain, mainly MV switch, Infinity switch,
[03:47] mainly MV switch, Infinity switch, um
[03:49] um there is an option of a scaling up which
[03:51] there is an option of a scaling up which is using the same strategy that uh we
[03:54] is using the same strategy that uh we have been seeing on the scale out, which
[03:55] have been seeing on the scale out, which is putting actual optical switches to
[03:58] is putting actual optical switches to interconnect these electronic
[04:00] interconnect these electronic uh high-capacity switches for the for
[04:02] uh high-capacity switches for the for that domain.
[04:04] that domain. On the bottom side, you you see the the
[04:06] On the bottom side, you you see the the requirements that we have been hearing requirements that we have been hearing from the industries.
[04:09] For that to really happen, we will need an extreme miniaturization um of of these OCS optical circuit switches.
[04:20] Industry's even talking about four OCS per one U rack.
[04:23] We need to be almost invisible [snorts] for next-generation clusters in order to to actually make this happen.
[04:30] We would also need to have an extreme cost-effectiveness.
[04:34] $100 per port has been a number that has been consistent over the last 3 years as a key target.
[04:41] And the reconfiguration time is not really a game-changer, but should be around the millisecond range.
[04:46] The options that are today are mainly relying on either 3D MEMS, are relying on piezoelectrics, are relying on on liquid crystal.
[05:00] Here you see some examples from Lumentum, from Coherent, uh from companies in China like Trellis Done, Polatis, and so on.
[05:07] Polatis, and so on.
[05:09] And uh they are great for the for for for the.
[05:12] they are great for the for for for the scale out applications or applications.
[05:15] scale out applications or applications not requiring extreme miniaturization.
[05:17] not requiring extreme miniaturization.
[05:19] And the good news is that they have been there for more than 20 25 years as they.
[05:22] there for more than 20 25 years as they play a key role in the telecom uh deployments.
[05:26] deployments.
[05:29] At the present, we are trying to to get the the metrics right for applications that demand extreme miniaturization,
[05:33] that demand extreme miniaturization, extreme cost-effectiveness.
[05:36] And that's why we are relying on silicon photonics.
[05:39] why we are relying on silicon photonics.
[05:41] What you see here is one graph that uh my peer Luis and I were obsessed with,
[05:44] my peer Luis and I were obsessed with, which is seeing the growth of photonics as an.
[05:46] which is seeing the growth of photonics as an integrated programmable photonics in.
[05:48] integrated programmable photonics in particular as a technology that actually allows you to use high volume.
[05:50] particular as a technology that actually allows you to use high volume manufacturing.
[05:55] manufacturing to actually exploit the power of integrated optics.
[05:56] to actually exploit the power of integrated optics.
[06:00] integrated optics.
[06:01] If you are familiar with transceivers, typically you find dozens of components, mainly high-speed modulators, high-speed photodetectors, but the technology we.
[06:09] photodetectors, but the technology we are talking here is a large-scale silicon photonics allowing.
[06:13] large-scale silicon photonics allowing the integration of thousands of actuators you see in the top bar like a 2 by 2 Mach-Zehnder interferometer, a typical 2 by 2 switch extractor.
[06:20] 2 by 2 Mach-Zehnder interferometer, a typical 2 by 2 switch extractor.
[06:23] If you are able to create fabrics that integrates thousands of these elements in a very clever way, you can actually get the same optical switching performance that you typically get with MEMS, mirrors, collimators and lenses, but now in a very tiny form factor.
[06:35] MEMS, mirrors, collimators and lenses, but now in a very tiny form factor.
[06:38] Very tiny meaning 20 by 20 mm chips that you can actually use now use similar process flows as the compact catch optics industry is now getting into the data center, but in this case with fully optical input, fully optical outputs, and switching in the optical domain.
[06:56] optical input, fully optical outputs, and switching in the optical domain.
[06:59] To do that, we have developed across these years a technology starts stack that takes a reliable photonic integrated circuits components.
[07:06] that takes a reliable photonic integrated circuits components.
[07:08] I mentioned before the Mach-Zehnder
[07:10] mentioned before the Mach-Zehnder interferometer.
[07:11] interferometer.
[07:13] We are talking here about ways to couple the light from the optical fiber right to the photonic integrated circuit, splitters, combiners, phase shifters, crossings.
[07:20] crossings.
[07:24] And at AtaiTecs we we are doing that.
[07:26] Let me say really well, but with a standard silicon photonics.
[07:29] So, we have been using many foundries across the world and replicating our process and getting beyond the state of the art performance thanks to to the photonic design, but again with a standard silicon photonics.
[07:41] silicon photonics.
[07:43] Key part, no movable parts.
[07:46] So, we don't have uh mirrors that are moving to actually employ the switching functionality.
[07:51] We are employing the the interferometric behavior of the light, so no movable part increasing reliability.
[07:57] When you combine these um fabrics advanced fabrics with the the control electronics, a software layer, you can create a fully functional device.
[08:03] We've been shipping that uh since the last year OFC, around April.
[08:10] since the last year OFC, around April.
[08:12] to the key hyperscalers, the key players
[08:13] the key hyperscalers, the key players that are building next generation AI
[08:15] infrastructure, and we are now in the
[08:18] infrastructure, and we are now in the evaluation phase on on more complex
[08:20] network tests.
[08:22] network tests. I wanted to show a
[08:24] I wanted to show a When I refer to key building blocks and
[08:27] that we are getting outstanding
[08:29] performance, this is just a
[08:31] a quick view of a 3D splitter
[08:34] and uh
[08:35] a monolayer crossing. So, you can see
[08:38] how for different generations across
[08:40] these 10 years of developments, we have
[08:42] been mastering standard processes and
[08:45] getting manufacturable
[08:47] fab tolerant devices.
[08:49] fab tolerant devices. How this translates when you put all
[08:51] together on on a photonic integrated
[08:53] circuit,
[08:55] you see here the
[08:56] what we consider state of the art uh
[08:59] designs that we have manufactured and
[09:01] test.
[09:02] 32 by 32 optical switch, our first
[09:04] generation. The next generation that we
[09:07] are packaging as we speak, and our 64 by
[09:10] 64 die.
[09:11] 64 die. All of them are 20 by 23 mm chips.
[09:15] All of them are 20 by 23 mm chips.
[09:18] And um thanks to these, you can think it now about the miniaturization and the compactness.
[09:21] I saw before prior slides with 2U racks, 3U racks, 8U racks.
[09:27] We are introducing these tiny chips, multiplying the radics every single year to target the the radics requirement for next generation scale-up networks, and we are embedding these in a ultra miniaturized form factor.
[09:41] In this case, we call it optical system-on-a-module.
[09:44] It's a device that actually is more or less this size.
[09:48] You can hold with a with a single hand, and then you squeeze that in in a in in the socket, and it's automatically connects to the back panel.
[09:58] So, basically, is a the the external fibers are just for test purposes in this case, and what we are showing here is an optical multiplane.
[10:07] So, we are able to feed with this design I'm showing here around three optical circuit switches per one U rack.
[10:09] Is this enough?
[10:11] It's way
[10:14] per one U rack.
[10:14] Is this enough?
[10:18] It's way beyond the state of the art today um by beyond the state of the art today um by more than one order of magnitude.
[10:19] It's more than one order of magnitude.
[10:22] It's great for the first uh great for the first uh POCs and proof of works,
[10:24] POCs and proof of works, but uh we believe that we can multiply
[10:26] but uh we believe that we can multiply this number
[10:28] this number almost by by 4x uh getting uh this
[10:32] almost by by 4x uh getting uh this invisibility in next generation networks
[10:35] invisibility in next generation networks uh as I was saying before.
[10:36] uh as I was saying before.
[10:41] What that means from um an application level point is if you
[10:43] an application level point is if you imagine, and this this was public
[10:45] imagine, and this this was public recently in a post if you are trying to
[10:47] recently in a post if you are trying to build a 576 uh GPU cluster,
[10:52] build a 576 uh GPU cluster, and uh you aim to
[10:55] and uh you aim to >> [clears throat] >> to wire this up
[10:55] >> [clears throat] >> to wire this up
[10:58] >> to wire this up in the in the scale-up domain,
[11:00] in the in the scale-up domain, uh the difference between having silicon
[11:03] uh the difference between having silicon photonics and using a alternative
[11:05] photonics and using a alternative technologies directly translates on the
[11:07] technologies directly translates on the footprint in the data center, which is
[11:09] footprint in the data center, which is quite expensive.
[11:12] We are talking about at least um requiring in this case 14
[11:15] at least um requiring in this case 14 cabinets.
[11:17] cabinets just for the allocation of optical.
[11:19] just for the allocation of optical circuit switches versus being able to.
[11:22] circuit switches versus being able to integrate the optical circuit switches.
[11:23] integrate the optical circuit switches within or just immediately next to.
[11:26] within or just immediately next to almost invisibly uh the the actual.
[11:28] almost invisibly uh the the actual computing and the actual networking that.
[11:32] computing and the actual networking that is being deployed today, the the regular.
[11:34] is being deployed today, the the regular scale out networks.
[11:36] scale out networks.
[11:37] So, that's.
[11:39] This is a must-have for next generation deployments, and we believe we have the.
[11:42] deployments, and we believe we have the technology to actually enable these next.
[11:44] technology to actually enable these next generation designs.
[11:47] generation designs.
[11:50] I would like to finish just touching a little bit what this implies in in terms.
[11:53] little bit what this implies in in terms of.
[11:54] of performance.
[11:57] In this case, the regarding energy and cost, the up the contribution.
[12:00] energy and cost, the up the contribution of the OCS would be also minimal.
[12:04] Here in this graphs, we are only considering.
[12:05] in this graphs, we are only considering the networking aspect.
[12:08] So, these numbers are even smaller if you consider the.
[12:10] are even smaller if you consider the networking and the computing.
[12:12] We are talking, according to our estimations.
[12:14] talking, according to our estimations, that only 3% of the total power.
[12:16] that only 3% of the total power consumption would be coming from the optical circuit switches.
[12:20] This is thanks to the very low power operation and the removal of a big portion of transceivers versus electronic high-capacity switches.
[12:30] Regarding another bit will be the reconfiguration speed.
[12:34] Today, the requirement in most cases is 50 milliseconds.
[12:38] 50 milliseconds, hundreds of milliseconds is something that is achievable today by the existing first generation optical circuit switches.
[12:47] I'm referring to MEMS and other technologies.
[12:49] So, this is not a particular specification that requires a lot of research.
[12:54] The good news is that the silicon photonics switching technology I presented here today has the capability to go 1,000 times faster.
[13:03] What that this means is is not an immediate advantage for the applications that are envisioned today, but there is already a lot of research from many companies and institutions around the world that are actually proving that
[13:18] world that are actually proving that being faster allows you to increase the granularity of or how frequent you actually reconfigure the network in order to maximize performance.
[13:29] This means that if you are able to go rather than on the hundreds of millisecond reconfiguration time to go down to sub 1 millisecond, then you have the opportunity to start doing reconfiguration at the intra job levels within the actual primitive communication libraries, and this provides a an opportunity to to optimize the network even with higher granularity.
[13:53] So, we would expect this to be a um an interesting competitive advantage in the future.
[13:59] I want to I would like to finish with just a quick um glance on on the device that we have the engineering sample that we have been putting in the customer hands for a year now.
[14:12] You see the transparency uh around the the wavelengths of operation of the data center.
[14:17] In this particular is the optical network engine 32 by 32.
[14:20] network engine 32 by 32.
[14:22] We are representing the typical four wavelengths of FR range, also the ones
[14:24] of the L WDM, DR,
[14:27] and you see that the performance of this one in particular is around plus minus 1 dB.
[14:31] Why is this?
[14:33] Because we are combining silicon photonics with inline amplification in order to compensate for that.
[14:37] So, we provide a lossless or gain control device.
[14:43] So, final call to action.
[14:46] Our products is participating and leading one of the groups sub groups of the networking area at OCP.
[14:54] In particular, we are collating with Lumentum an action on optical circuit switches.
[14:58] So, you are welcome to join and to see the progress of this project.
[15:03] Thank you.
[15:16] Can I just ask if it's silicon photonics, does that mean it is primarily for scale up, and do you
[15:22] primarily for scale up, and do you foresee a time when it will be large?
[15:25] foresee a time when it will be large enough to compete with the incumbent MEMS?
[15:29] MEMS? Yes, so um definitely uh specifications for scale-up are um are a match for silicon photonics.
[15:39] Uh the for the other segments, a scale-out, a scale-across, we in particular have a a way to use silicon photonics to increase that to we call it infinity radix, like very high radix approach.
[15:54] So, as we see the market is up to 144, will be dominated by silicon photonics.
[15:59] Then from 144 to around 500, maybe dominated by MEMS and other technologies.
[16:06] And then from 500s and beyond, and we have people asking for this type of technology, will be again dominated by silicon photonics or co-share with MEMS-based devices.
[16:17] So, I think it's going to be a heterogeneous. And very briefly, is there an issue with all the number of components or
[16:23] all the number of components or elements, photonic elements in this as elements, photonic elements in this as you scale-up?
[16:25] And silicon photonics can be lossy if there's a large number, so people talk about adding gain, but that's more componentry.
[16:28] So, how do you tackle that?
[16:33] So, we are actually introducing gain in-line amplification.
[16:37] Uh today we have around 10 dB loss.
[16:40] We compensate that with 10 dB gain.
[16:42] It's a gain which is quite small or low compared to the typical gains employed in telecom, which is the 25 dB or or and beyond.
[16:47] So, yes, today we increase we put the gain to have a gain control device, but the way we are moving forward is we predict that we are going to be able to reach around 4 dB loss even increasing the radix.
[16:51] And this is thanks to the innovation we have done at the fabric and our an internal chip architecture level.
[16:58] So, our path is moving to 4 dB loss approximately, and and the same time having the option of uh incorporating or not the amplifier for different applications.
[17:01] Thank you very much.
[17:27] Can you comment on possibly other wavelength bands?
[17:31] I saw you had the 1310, any work on C-band or any work down in the LED ranges?
[17:35] So, in principle, the transition to full O-band coverage to C-band or C plus L should be relatively straightforward.
[17:42] We are focused on the intra data center opportunity.
[17:48] Having a massive O plus C plus L, it's something that with silicon photonics, it's uh it's really hard to get.
[17:55] There is There are ways to get there, but uh we are not focused on that.
[18:00] So, you could have a specialized O-band or a specialized C plus L band devices for the telecom and datacom applications.
[18:05] you put up that power slide real quick again?
[18:09] It went by real fast.
[18:10] Uh I I I don't think I have the control of that, but uh So, it Can you comment on the compare between I saw it was less than 3% of the power for the optical switch, but what is it today?
[18:22] Like what Compared to today, how much does switching take of the power budget?
[18:24] Cuz it should be a huge savings in power, right?
[18:25] Uh in in the case I was putting
[18:28] right?
[18:28] Uh in in the case I was putting on the slide, not really, because I was on the slide, not really, because I was mainly talking about the case where you,
[18:31] mainly talking about the case where you, rather than imagine that you are rather than imagine that you are substituting the intra cabinet switches by an optical one, what we are doing and by an optical one, what we are doing and the industry's moving towards is the industry's moving towards is basically multiplying the number of ports that you interconnect together,
[18:44] ports that you interconnect together, and you still claim that this is a scale-up domain.
[18:46] So, you still have electronic switches per port, and then you are using optical switches to actually expand the scale-up domain in number of GPUs.
[18:54] Thank you.
[19:00] My last question.
[19:02] Okay.
[19:02] Uh you're sending very high-speed data through all it's agnostic to data rate.
[19:06] Yes.
[19:06] You're sending it through a train of Maxender from and gain and gain switches.
[19:12] Does this affect the uh signal integrity at all?
[19:14] Yes, so beta rate performance is one of the metrics that we monitor.
[19:20] Um the the impairments on the beta rates may be coming from different sources.
[19:24] It can be crosstalk, it can be from the loss per sec.
[19:27] In our case and and most
[19:30] Loss per sec.
[19:30] In our case and and most of the concern is on the gain control of the concern is on the gain control devices.
[19:35] The gain typically produce a penalty on the BR.
[19:38] In our case, we are keeping that under control with around one decade, one decade and a half degradation for many formats and many speeds.
[19:47] And that's only possible if you design the moderate gain amplifier to feature low noise figure and very good saturation power.
[19:56] So it's possible to get minimal BR penalty degradation.
[19:59] Okay, thank you very much.
[20:00] Let's thank our speaker.
[20:02] >> [applause]