# Optica Executive Forum at OFC 2026. March 16th 2026 Panel: Scale Out Session

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

[00:00] We would like now to continue with the with the program and for that I would like to bring to the floor our next member of the optica executive forum organizing committee.
[00:10] He is the chief marketing officer of coherent Sanjay welcome to the stage thank you very much what you have done.
[00:23] Thanks Jose.
[00:23] Hi I'm Sanjay I am the CMO of coherent.
[00:27] I'm super excited to be here again and it's hard to see the faces but I've it's nice to see so many familiar faces earlier uh at the break and we've got we've got a really amazing panel for you guys today and where is VJ VJ said eighth time I'm going to do it for the ninth time so as a committee this year you know we decided to organize the panels by one two three right scale out scale up and scale across and Tim started us with uh with scale across and with scale across and
[01:01] now it's scale out scaleout is the economic center of AI connectivity today.
[01:07] right we are seeing um you know a fantastic epic demand of of uh of across the entire patch across the entire supply chain.
[01:16] the scaleout networks are kind of characterized by sort of a broad interoperable ecosystem a standardized broad interoperable ecosystem.
[01:27] And it's been a relentless focus on again three things.
[01:29] Bandwidth, power, and cost.
[01:34] And um to help us unravel all of that today, we got an illustrious group of uh speakers.
[01:39] And they don't need any introductions, but I will I will introduce them very briefly.
[01:47] Um the other the other thing in with scaleout networks today it's you know pluggables from a technology perspective it's it's all mostly pluggables today but with CPO and NPO you know on the horizon emerging very rapidly and it's not about just
[02:02] delivering the technology it's the ability to do it in scale and in a scale that we have not seen before and and boy are we all getting ready for it so it's it's pretty exciting times so I don't know if Near is here.
[02:15] Oh, okay.
[02:15] Cool.
[02:19] We were Okay, so we'll start with uh well, Near Margalade from Broadcom will start us out first.
[02:24] Uh Near is VP of their optical systems division.
[02:27] Then N will pass it on to Rada Nagarajin who's a CTO VP of Marll and then Rada will pass it on to Whipplebud who runs corporate strategic marketing at Coherent and then Whipple will then pass it on to Ryan Yu who's VP of marketing at Terraop.
[02:44] Okay, with that let me let me ask invite Ne to the stage.
[02:47] Thank you.
[02:51] Thank you.
[02:54] Thank you so much for inviting us.
[02:55] All right.
[02:57] So, we'll talk a little bit uh about uh uh scale out uh and what we're doing there and also
[03:03] we'll talk a little bit about the interplay of scale out and scale up.
[03:06] So, that uh should be interesting.
[03:08] So, one of the first questions that comes is people talk all the time on 1 million GPU or XPU clusters and how are you going to do that.
[03:18] So, uh we'd like to build it in the most efficient way possible and you can't have switch radixes.
[03:24] We're making the assumption that they're currently limited to, you know, in production right now, maybe 512, maybe 1024 type switch radiuses.
[03:33] And if you just do the math on that and use traditional clawbased networks, uh you need several layers, two three layers, and you need uh uh optical transceivers and a connectivity and you burn a ton of power uh providing that connectivity.
[03:50] So is there a better solution uh to doing that?
[03:51] And the answer is there probably is.
[03:55] So uh one way to do that is to actually leverage your scale up network.
[04:01] Now people are
[04:03] going to have uh talks later on how to make your scale up network a larger domain and people are talking about getting to 1024 even larger uh basic layer.
[04:13] And what does the scale up domain uh really drive is very high bandwidth low latency uh interconnects sometimes with a different protocol sometimes with the Ethernet uh but Broadcom's been driving uh this standard uh called EON through a multissource agreement with a number of people that allows us to use the same messaging system for scale up and scale out systems.
[04:37] So you can actually treat uh your scaleup network as part of your scale out network.
[04:42] So rather than uh just treating each uh GPU as an independent node that has to communicate with all the other nodes, you can start thinking about the the node as a pod.
[04:55] Uh let's take an example of a thousand node uh pod.
[05:03] a million XPUs, how many uh pods do I need?
[05:06] Well, I need a thousand a thousand uh pods.
[05:09] And what that means is my single layer Ethernet switch can connect all million uh all million uh GPUs, XPUs.
[05:19] Uh if you can take advantage of of this uh scaleup system.
[05:22] So that that's going to be uh the core of my talk and just talking about what what that looks like and uh what kind of connectivity you want to have in your scaleup network to enable that type of uh solution.
[05:37] So, uh, one of the things that we're going to focus on is there's a number of different optical technologies that can be used for, uh, connecting, uh, optically connecting.
[05:47] So, if you're using copper alone, you're probably going to be stretched to reach a thousand or larger scale up networks.
[05:53] You're going to need to use optical technology.
[05:55] And there's a number of optical technologies that are available.
[05:57] One is continuing kind of the ITE path which is kind of DRbased uh optics going
[06:05] from 100 gig and currently being deployed 200 gig and I think there's going to be many demos here at this show for 400 gig uh optical technology uh but it's going to be hard to go beyond that uh data rate and your power efficiency on those optics as they get uh higher speed uh are limited by the combination of you still need copper certis to reach those optics and the optics optics themselves drive a certain amount of power.
[06:30] So typically the optics even if you're in LPO mode might be 5 seven pjles per bit and then you're going to add another four pjles per bit for uh for the 30s to reach the optics.
[06:42] So the true IO power is almost 10 pigs per bit which is really not very uh very efficient for uh for scale up systems because we know in scale up bandwidth is 10 times higher than scale out the impact of excess power consumption is really really very severe and I think you'll see a number of different uh technologies going on the shows here that people have been working on uh
[07:05] large uh wavelength combs or what's called slow and wide where you run many wavelength uh together uh to get your uh get your power consumption uh lower.
[07:18] Of course, if you go to too many wavelengths, uh it starts being more complicated in terms of laser sources.
[07:25] You need special specialty laser sources and specialty controls and ring structures.
[07:30] So there there are some challenges uh with those.
[07:32] And I've heard a lot uh in previous shows uh why not do both fast and wide and uh everyone's like oh that seems like a great idea right why not why not go 16 wavelengths at 400 gig right that sounds you know the best of everything right and many people have advocated this uh the problem uh with that is that uh what that translates if I want to use a thousand radex switch and I each of my ports herz is 6.4 terabits uh per
[08:06] second. I need a 6.6 pedabit switch.
[08:10] And you guys know in the market that Broadcom just released a 100 terabit switch.
[08:14] So 64 times larger than what's available in the market.
[08:17] So that's not really a good fit for what's available in the silicon.
[08:21] And one of the key messages that I'll say in the talk is the switch and the optics have to be married together and they have to match.
[08:28] can't get one ahead of the other one in in terms of uh if you want to maintain radics and you want to maintain bandwidth.
[08:36] So fast wide not really a realistic solution for many years uh out in the future.
[08:45] What about uh really wide so microledds uh lots of uh 10 gigabit or 20 gigabit uh signals uh you can certainly get power consumption pretty low uh on that systems.
[08:57] But if you think about how these scaleup systems are interconnected, every GPU port has to be connected to every switch port to be able to give you all to all bandwidth.
[09:08] So fiber becomes a really uh big issue.
[09:11] And if you're going really slow, really wide, you're dealing with a lot of fibers.
[09:15] And not just pointto-point fibers, you've got to uh what's called shuffle all the fibers together.
[09:18] So all your interconnects, your optical connectors.
[09:22] Uh it's not just a pointto-point connection that you can do with uh LEDs.
[09:25] If you're going to do a true scaleup network, you got to be able to do optical connectors.
[09:29] You got to be able to shuffle all these fibers.
[09:31] So, uh, fiber becomes start a really really limiting factor if you're going, uh, too slow.
[09:37] So, what do you do?
[09:41] Uh, so one obvious thing that's a really easy win that doesn't need a huge amount of work is just do both directions on the same fiber and that's just called BA.
[09:48] It's been used in fiber the home for I don't know 20 years, 10 years.
[09:53] uh and we've we've done that technology.
[09:56] It works uh uh really well uh and it does not impact your radex but it reduces your fiber count.
[10:04] So that's that's and that's that's a really uh really easy win.
[10:06] What
[10:09] Else do you do?
[10:12] Uh you have to match the bandwidth per fiber uh to uh the switch sizes that you're able to build.
[10:19] So, if you're at a 100 terabyt uh switch, you probably don't want to go more than 200 gigabits per fiber.
[10:26] And if you're at 200, you probably don't want to exceed 400 and uh and so forth.
[10:30] So, there's plus or minus a factor or two on these things, but you know, this is this is a relatively good rule of thumb is these two fiber bandwidth and switch bandwidth.
[10:41] uh if you're doing an alltoall connectivity uh you need to marry them in a way that still preserves your radex.
[10:49] If you use too much bandwidth per fiber you're consuming out your radex and your radex becomes smaller uh if you're if you're using too small too small a bandwidth per fiber you're going to get hit with uh with fiber cost.
[11:04] So uh to that effect uh we announced just recently with a
[11:10] number of key industry partners the OCI MSA that really tries to match all these different factors uh together getting the right switch size the right bandwidth per fiber using by technology not not needing advanced uh laser comb technology we can use regular DFB lasers uh and use low speed NRZ gets the power extremely uh low and allows for the scale up network which is uh which is a key enabler to a really uh very high rail count scale out network.
[11:41] So if you can get your scale up network to be Ethernet based and be able to route across that to be able to access all the different nodes of your scaleout network you can really limit your the size and uh number of layers that you have to do in a scaleup.
[11:58] I think that's uh it for this one.
[12:01] We'll pass it on to Rada.
[12:04] Um I'm Rada Nagaraj and uh I I deal with um optical engineering at Marvel.
[12:07] All right, we'll we'll get started.
[12:14] Oh, uh, this is something my legal wants you to read.
[12:16] If we can actually go through the whole thing, man, you're good.
[12:20] Um, so let's get to the meat of the presentation here.
[12:23] Uh, we've been talking about um scale up.
[12:26] We also have a scale in uh up and out and across.
[12:35] The point is um these are just demarcations, shorthand notations to um hopefully demarcate the technologies in between.
[12:43] So there are a couple of things that are happening here, right?
[12:45] Because we've seen this so often, uh, by the time you're out of this, you you'll be tired.
[12:50] There are two things that are happening.
[12:52] Um, one is because we're talking about scale out, which is essentially sort of within the data center.
[13:01] The way I like to say is the coherent is moving in.
[13:05] And uh, hopefully the co-package technologies developed for within rack interconnects are moving out.
[13:11] So the
[13:15] squeeze is somewhere in the middle um on the scale out networks but scale out is is the hot is the engine that fuels this industry.
[13:24] uh pluggables completely dominate and from the right hand side um it's it's coherent moving in so it's pluggables and pluggables and um on the left hand side scale up moving and it's mostly to do with form factor um higher bandwidth density greater levels of integration.
[13:46] so let's focus on each one of them I I'm going to start on the on the left and um sort of move my way uh to the right Um,
[13:55] I'm talking about my right and left.
[13:56] I don't know what it looks like on the slides.
[14:00] So, all right, let's let's going to get a little bit more into the technology side.
[14:07] Um, this is sort of the light engine architecture.
[14:09] Uh, it's one of many.
[14:12] The reason I bring this up is this is more
[14:15] than about co-packaging.
[14:18] You know, when you look at a cross-section like this, we go, "Oh god, it's co-packaging."
[14:21] No, it's got to do with speed.
[14:23] has got to do with integration.
[14:25] It's got to do with signal integrity.
[14:27] Uh this particular architecture has got a top layer of electronics integrated onto a photonix layer and all the little gaps that you see in the photonics layer is TSV based goes in the substrate and integrated the back of the substrate.
[14:40] This has become sort of the vehicle for us to do almost anything uh these days because wire bonds don't cut it.
[14:49] And one example um is a light engine on the uh left hand side here.
[14:55] Uh this is a 6.4t light engine 32 lanes of 224 gig.
[15:00] And you could use a light engine uh not necessarily for scale up for scale out uh from XPUs as well.
[15:05] I mean there's always this belief that XPU is always going to a scale up network.
[15:09] Not necessarily.
[15:13] uh you could do a scale out network as well and you can use the
[15:17] light engine into pluggables
[15:19] and and you can use these compact form
[15:24] of uh light engines uh people are exploring a lot of near packaged optics
[15:29] uh for scale up and scale out as well
[15:32] but it's mostly the scale out I'm focusing and scale out usually runs in this Ethernet u
[15:37] sort of data rates we'll come to that in a moment
[15:44] And of course you can use the same light engines for a switch interconnect.
[15:49] There was a lot of concernation about cooling uh the light engines together with a switch.
[15:54] But these days we've gone past that.
[15:57] Almost everything that we have today is liquid cooled including uh XPOS's these days are pluggable liquid cooled.
[16:04] So, we got to get past this air cooled infrastructure and really get to an liquid cooled infrastructure to realize any of this.
[16:11] We're not pretending that the that thermal densities are are manageable with air
[16:17] cooling anymore, right?
[16:22] And now moving to the right hand side,
[16:24] moving in uh coherent light leverages on the uh scale out or the DCI type interconnects.
[16:34] Uh we have two generations of field proven.
[16:36] They're not all coherent by the way.
[16:38] Uh the asterisk says our first generation colors was actually a PAM 4 interconnect but it went um 80 km.
[16:47] The interesting thing to watch between these generations is the power consumption.
[16:49] I mean start at 4 1/2 watts per 100 gig and sort of stayed at 4 and a half four and a half watts at 800 went down to 3 and a half and at 1.60.
[17:00] It's in design stages.
[17:03] these days about two watts per 100 gig.
[17:06] But the important point is it points to one of the charts that was in the previous session.
[17:09] There's a convergence.
[17:12] Uh this is a ZR module in power consumption looking very close to a 1.6
[17:20] PAM 4.
[17:22] In a coherent light type module, the convergence would would even uh it'll be even more um apparent.
[17:31] So coherence isn't for these large boxes interconnecting as was already shown very well uh in the first scale uh scale across session.
[17:39] So if you leverage this if you take this in I think before that we're going to go into the technology piece first.
[17:47] Uh we talk about uh well let's talk about the uh the the reach coming in.
[17:51] So where where does coherent light play?
[17:56] Um again the previous session was a great setup that they want coherent they would like coherent light.
[18:03] The real battle we think is or everybody thinks is going to happen is 3.2T 2T optics at 400 gig per lane.
[18:09] At 200 gig per lane today there are some small deployments.
[18:14] Uh there was a demonstration last year OFC coherent
[18:20] light but that was at 100 gig and 200 gig.
[18:23] uh DSPs are are are uh coming into play.
[18:27] At 200 gig there will be some deployments.
[18:29] Uh it's not a it's not when it's going to come in.
[18:32] It's it's uh it's not an if it's when it's going to come in.
[18:34] At 400 gig, there are some real challenges with IMDD.
[18:36] I'm not saying IMDD is going away, by the way.
[18:39] Ease of use and scale of deployment, you know, fighting incumbency.
[18:41] That's not the point.
[18:43] The point is to be able to complement the technology and have core deployments at 400 gig.
[18:46] I mean, as dispersion scales at a square of the baud rate, uh, you need something else.
[18:51] And where the yellow box stops, you know, it could it could be a religious war whether it's at 500 meters or 300 or 1.1 kilometer.
[18:53] It doesn't matter.
[18:55] Uh scale out will see coherent light and that that's sort of a definitive statement here.
[18:57] You can take me to task over it.
[19:00] And uh it will it will be
[19:20] Optimized. I'll get to that in a moment.
[19:23] And before that, uh, why is is is is the question always is?
[19:30] Would we ever get the coherent light DSP powers low enough that we will deploy?
[19:35] The answer is yeah, we will uh because there are a number of differences between ZR, ZR plus column and coherent light.
[19:44] Although this is not quantitative, you know, you could argue over 3 plus over two.
[19:47] The point is if you're familiar with the FEC inside uh data centers, it's not as as as what should I say, power hungry or robust as ZR ZR plus bulk chromatic dispersion.
[19:58] Uh, famous last words because that 448 gig is is 4x over 200.
[20:05] We might need quite a bit of dispersion compensation.
[20:09] We don't have nonlinearity compensation.
[20:11] Probably it's a constellation shaping.
[20:13] We don't have to use a tunable laser.
[20:15] DFBs work really well and strict line with requirements and in the end, they're both
[20:22] OSF modules anyway.
[20:25] So pluggable sa pluggable and power
[20:28] consumption will depend on the baud rate
[20:29] the number of lambda lambdas and how
[20:31] they scale. So there is a path and being
[20:35] actively explored
[20:37] uh so that this power consumption was
[20:39] somewhere in between IMD and um and uh
[20:44] the real coherent
[20:47] and
[20:50] are we anywhere uh ready for 400 gig uh
[20:54] generation? The answer is absolutely
[20:56] yes. Uh I thought I would never see this
[20:59] or well retired by the time I saw any
[21:02] analog components with over 100
[21:03] gigahertz bandwidth. Um apparently it's
[21:07] it's it's everybody does it every day
[21:09] these days. Uh the first chart shows the
[21:12] optical modulator bandwidth for us. Four
[21:14] generations on the chart that I showed
[21:16] you went from 100 200 100 400 800600
[21:20] those shows the first four generation of
[21:22] modulator bandwidth. The next are photo
[21:25] detectors. You need modulators and photo
[21:26] detectors. By the way, the focus always
[21:28] seems to be on modulators and
[21:30] electronics. We do electronics really
[21:32] well as well. You need all three pieces
[21:35] and we have electronics um that that
[21:37] measurement stops at 145 because the the
[21:40] all equipment related are limited except
[21:42] for the um except for the optical
[21:44] detector bandwidth.
[21:48] Thank you.
[21:52] Good morning and thank you for your
[21:54] attention.
[21:56] Let me raise two talking points which
[21:58] then we can lead to discussion for. As
[22:03] Ethernet optics began to become
[22:06] applicable to AI over the last couple of
[22:09] years there are two somewhat obsolete
[22:13] assumptions that people operate with.
[22:15] One is for scale out it's the same
[22:18] optics not quite and two for growth in
[22:22] optics when we go to scale up we will
[22:24] use the same optics maybe maybe not let
[22:28] me set the stage the first thing to
[22:31] remember is that scale up and scale out
[22:34] support different system objectives
[22:37] and it really boils down to what at the
[22:40] end of my three slides uh you'll see as
[22:43] an interconnect contract as I call it.
[22:46] Are we prioritizing latency determinism
[22:50] or an engineered tolerance of a link?
[22:53] This will become more clear as I go
[22:55] further. For one thing,
[22:58] scale up optics and scale out optics can
[23:01] evolve separately because the scale up
[23:05] and scale out are architecturally
[23:07] separate. They can have different lane
[23:10] speeds. When you implement optics, they
[23:12] will definitely have different aggregate
[23:13] bandwidth.
[23:15] When you go to scale up, you're going to
[23:17] need what we call coherency class
[23:19] optics. That is it. It must feel
[23:22] synchronous, very low error and an
[23:26] extremely tight limit on latency
[23:28] variation, both static latency and
[23:32] dynamic latency that ensues from let's
[23:35] say errors and retransmissions.
[23:39] Scale out on the other hand is more
[23:40] tolerant because it's going across a
[23:44] larger geographic span and the links
[23:46] need to be able to give you some errors
[23:50] because in return you have a contract
[23:52] with the system architect which is that
[23:54] they will tolerate a limited amount of
[23:57] errors and delay in an engineered way.
[24:01] It isn't the same as your grandfather's
[24:05] ethernet optics of 2020. it is more
[24:08] strict. It does demand lower error rates
[24:13] and so you do take higher bandwidth
[24:15] components and inject find ways to do uh
[24:17] give it higher signal to noise ratio.
[24:20] This distinction is not hard. But as we
[24:23] go to future or futuristic physical
[24:27] layers, put up satellites in space or
[24:30] put data centers in the ocean.
[24:33] Let's just take those two examples. I
[24:35] would say given that background perhaps
[24:39] a cluster of satellites is more scale
[24:42] out because it has to have engineered
[24:44] tolerance of link failures while the
[24:48] ocean pipes can be deterministic and
[24:51] therefore it could be a scale up cluster
[24:53] as long as it's tight.
[24:56] But as this illustrates how soft uh this
[25:00] distinction is.
[25:02] One thing that the legacy has inherited
[25:05] successfully from Ethernet is that
[25:08] scaleout interconnects
[25:11] have successfully operated when they
[25:13] have combined three system design
[25:15] objectives.
[25:17] Performance, interoperability and
[25:20] pluggability.
[25:22] On performance, sure every generation
[25:24] you go higher speed. At higher speed you
[25:26] have to deal with your signal amplitude,
[25:30] receiver sensitivity, errors, noise and
[25:32] so on.
[25:34] Luckily the components are keeping up.
[25:36] So from 100 gig to 200 gig at 400 gig we
[25:40] foresee indium phosphite modulators,
[25:43] differential EML components, high-speed
[25:45] photo diodes, silicon photonics with uh
[25:49] CW lasers uh and 200 gig pixels. So
[25:52] they're all stepping up to say you want
[25:55] to build a link the components are
[25:56] available
[25:58] but we still have to do the other two.
[26:00] One is interoperability
[26:02] unlike scale up where you can have a
[26:05] closed system in theory
[26:08] for scale out you must have a multi-
[26:10] vendor ecosystem that's an absolute must
[26:14] uh and therefore the standards and the
[26:16] MSAs. And may I add that this applies to
[26:19] CPO as well. If you are ever going to
[26:21] use CPO in scale out,
[26:25] I think it should begin with being on a
[26:27] socket.
[26:29] So on pluggability, absolutely
[26:31] preferred.
[26:33] We know that the front panel pluggables
[26:35] are thriving. You can see here these are
[26:37] the units growth uh by signal AI
[26:40] calculations.
[26:42] Um we've never seen a growth rate of
[26:44] 300% in six years ever before.
[26:48] The next switch which we expect to be
[26:50] a,000 IO 200 gig as Near mentioned uh
[26:55] will continue this tradition and to step
[26:59] up to keep up with that high radics we
[27:02] now have one option at least the XPO
[27:04] higher density form factor and for those
[27:07] who want to do CPO co-ackaged optics in
[27:12] scale out there is now an option for a
[27:15] new MSA a new form factor called open
[27:18] CPM X. It's not front panel pluggable,
[27:21] it's board pluggable. So if you want to
[27:23] ask what the distinction is, the way I
[27:25] see it is when you have a front panel
[27:28] pluggable, you almost have a a division
[27:31] along the other side where when you're
[27:34] standing outside the box, you can do one
[27:38] late binding, meaning I don't have to
[27:40] decide what optics I plug in today.
[27:44] um interoperability that's multi-
[27:47] vendors can plug in and they can
[27:49] interoperate and serviceability
[27:53] when you go to this open CPX form factor
[27:56] you probably
[27:58] don't make it as easy as before to have
[28:01] late binding because you have to open
[28:03] the box
[28:05] but it's doable and certainly you're
[28:08] retaining the benefits of multi- vendor
[28:10] ecosystem and serviceability because
[28:12] it's removable
[28:14] As my words suggest to you, maybe we
[28:17] need a new vocabulary to just not just
[28:20] say pluggable soldered
[28:24] for scale out. Definitely 400 gig per
[28:27] lane is a a very popular talking point.
[28:32] There are three types of 400 gig
[28:33] components that at least I am aware of
[28:36] uh that are keeping up with the
[28:38] requirement. One is a differential EML.
[28:41] differential as in if it's not
[28:43] single-ended you have signal integrity
[28:45] benefit second is believe it or not but
[28:48] a pure silicon PN junction mark sender
[28:52] modulator no TFLN no polymers um appears
[28:56] to be feasible you I welcome you're
[29:00] welcome to look at one of these post
[29:01] deadline papers on that and um pretty
[29:05] soon you will also see demonstrations
[29:07] and more physibility of that and then
[29:11] the indium phosphite modulator can sit
[29:14] on silicon and and we have a little
[29:17] video playing so you'll see that all
[29:20] three um options are technically
[29:23] feasible.
[29:25] The tradeoff at 400 is that you're not
[29:28] going to get the same reach. So the old
[29:31] Ethernet tradition of having a 2
[29:32] kilometer remains within a building
[29:35] which we really used for 4 dB loss
[29:38] budget and not there are no networks
[29:40] that are 2 km long within a building. Uh
[29:43] so we're going to have to be a little
[29:44] more nuanced about our vocabulary here
[29:47] as well. What may we may want to say is
[29:50] well the reach is never more than 500
[29:52] mters for a scale out network but we can
[29:54] still have a 4 dB loss budget or perhaps
[29:57] 3.5 and so on. So our vocabulary will
[30:01] evolve but these tradeoffs remain
[30:04] fundamental.
[30:06] At this point I actually had enough to
[30:09] summarize and give you uh a talking
[30:11] point for discussion but I was told that
[30:14] if I don't have a slide on CPO I'm not
[30:16] allowed to talk over here. So here's a
[30:19] slide on CPO.
[30:23] The only takeaway point that I'd like to
[30:25] leave you with is that for scale out, it
[30:29] might alter the mix of opportunity for
[30:32] optic suppliers. So if you are a
[30:33] pluggable supplier, you probably will
[30:36] broaden the scope of your portfolio uh
[30:39] to have these fiber harness and external
[30:42] laser source and products similar to
[30:44] that.
[30:45] But make no mistake, CPO whether in
[30:49] scale up or scale out is nothing but a
[30:52] bonanza for optics suppliers. It is an
[30:55] enormous market expansion opportunity.
[30:58] So good days are ahead of us no matter
[31:00] what you do in the industry.
[31:03] Let me recap. I'll focus on scale out
[31:06] because this panel is about scale out
[31:08] but I have scale up as reference just so
[31:11] we can do the contrast right
[31:14] in scale up you want to start with
[31:17] inherently high signal to noiseise ratio
[31:19] design which is why NRZ and slow and
[31:23] wide is an absolute option while in
[31:26] scale out you probably want a higher
[31:29] speed and so you take your existing
[31:31] components and infuse them with a higher
[31:34] signal to noise ratio. Um, but remember
[31:37] that when you go to 400 gig per lane,
[31:39] there will be trade-offs and therefore a
[31:41] lot of us are discussing the possibility
[31:45] of a very high radics 200 gig per lane
[31:48] system implementations and it's all good
[31:52] on pluggables. There's no doubt that the
[31:55] market is thriving and will continue to
[31:57] thrive. It's riding the experience
[31:59] curve. Thank you Ethernet. Thank you
[32:01] front-end networks.
[32:05] on interoperability. I again reiterate
[32:08] that unless we continue to have a strong
[32:12] version of it for whatever new speed and
[32:14] new design um we will not see as much
[32:18] success. So this is an absolute must
[32:20] that for scaleout optics a multi- vendor
[32:22] ecosystem must continue to thrive and
[32:26] I'm I'm very happy to see that the XPO
[32:28] form factor has stepped up as an example
[32:31] to have a higher density pluggable uh
[32:34] that continues to make these three
[32:35] objectives of scale out feasible
[32:38] while for scale up we may want a higher
[32:41] bandwidth density we may want a stronger
[32:44] consideration for CPO uh and as I said
[32:47] It's still um a very good um it's only
[32:51] good news for those who are in the
[32:53] business of supplying optics for that.
[32:55] It isn't one killing the other or one
[32:58] hurting the other. The expansion of the
[33:00] market is good for all of us. Thank you.
[33:07] >> Thank you all. Uh this is Ryan Yu from
[33:10] Terrahub. So we'll discuss uh
[33:14] how do we use uh open optics to scale AI
[33:18] infrastructures. So this is probably
[33:21] pass number seven now in terms of
[33:23] showing the scale up, scale out and
[33:26] scale across and also you've heard many
[33:30] uh previous speakers talking about the
[33:31] requirement optics. One thing I just
[33:34] like to emphasize is no matter what
[33:36] technology
[33:38] new shiny wares you're going to promote
[33:41] these days time to volume is really
[33:44] critical. Uh whatever you want to sell
[33:46] you want to promote at the end of day
[33:48] the end customer ramping up it's just so
[33:50] fast we want to make sure all the
[33:53] technologies we're promoting going to be
[33:55] hitting those key metrics but also you
[33:59] need to hit the volume very quickly. So
[34:01] that's actually really critical.
[34:04] Now I also like to talk about open
[34:07] optics actually really exemplified by
[34:10] the pluggable optics our industry have
[34:13] been enjoy in the last two decades. Uh
[34:16] these have serving us really well on AI
[34:18] infrastructure built out in the last few
[34:20] years as well. Uh here I just
[34:22] illustrating the 400 gig
[34:25] 800 gig now really ramping high volume.
[34:27] I think this year probably going to
[34:29] break record again and at the same time
[34:32] 1.6T is ramping. So you can see three
[34:35] generations previously already ramping
[34:38] really quickly and they are serving this
[34:40] very broad uh from the very short reach
[34:43] to middle reach to very long reach in
[34:45] the pluggable form factor uh well for us
[34:49] and obviously to go to the next
[34:51] generation we start to need to look at
[34:54] all these metrics as well to cover the
[34:57] whole AI infrastructure buildout needs.
[35:00] Uh one thing I just like to remind you
[35:02] uh no not all silicon photon are created
[35:05] equal. I I think you should make note
[35:08] that some of the silicon photon has been
[35:09] deployed in the last few years with very
[35:12] high volume. This is just our part of
[35:15] contribution. We actually ship over 15
[35:18] million silicon photon transceiver just
[35:20] for the last two years. And this number
[35:23] continue to climb very fast. And this is
[35:26] allow us to now build a very fundamental
[35:30] base for very high reliability,
[35:33] very good yield and operational capable
[35:37] to support high volume RAM when we need
[35:40] it.
[35:42] And another key factor is reliability is
[35:45] not just calculated. You actually need
[35:48] to deploy these optics, iron out the
[35:50] bugs particularly all the processes
[35:52] associated with building these parts.
[35:54] Then you can call it reliable. So this
[35:56] is again another contribution we've been
[35:58] making is build these parts in large
[36:00] volume and demonstrating they are very
[36:04] low failure rates.
[36:07] Now come to the new uh pluggable optics
[36:10] coming to the scene and many of you
[36:12] heard uh just before C uh there's XPO uh
[36:16] announced uh we are part of that
[36:18] co-founding team and then there's second
[36:21] we'll call it open CPXMSA is a pluggable
[36:25] on board but is a CPO application
[36:28] platforms so these two newer type of
[36:32] pluggable optics is allowing us to now
[36:35] expand and the capability both in
[36:38] capacity and density
[36:40] and also looking at integrating the
[36:43] liquid cooling into the whole pluggable
[36:46] form factor. So just looking at how we
[36:49] now going forward that we can see the S
[36:53] OSAP serving us so well in the three
[36:56] generations probably have another leg to
[36:58] go to 3.4T 40 right now I calling it
[37:01] OSAP plus because haven't really been
[37:03] officially launched as a new form factor
[37:06] but I think you know a few of us in the
[37:08] industry already started work on this as
[37:10] well and then SPO and open CPX MSA going
[37:14] to be coming in the pipeline so VJ
[37:19] actually talked about this earlier uh
[37:22] here I just like to say SPO is a new
[37:24] form factor actually supporting the
[37:26] whole spectrum of scale up applications
[37:29] and scale out and then scale across.
[37:33] Uh one thing nice about a uniform
[37:37] universal form factor is you can allow
[37:39] to incorporate many technologies
[37:42] prove it out in a pluggable form factor
[37:44] before actually hitting right next to
[37:46] the ASIC where it started become very
[37:48] expensive. So we do anticipate some of
[37:50] the newer technology can be included
[37:52] inside SPO form factor and start to
[37:55] offer deployment history and accumulate
[37:58] those knowledge and experience uh for
[38:00] these new type of uh optics coming in
[38:03] the pipeline.
[38:05] Uh here just to impress on you uh uh
[38:08] 12.8 SPL I think uh Andy probably show
[38:12] some of these slides yesterday. Um this
[38:15] is really giving you a very
[38:17] densification of switches. So this is
[38:19] very critical for scale up when the
[38:22] networking portion of the equipment now
[38:24] start to occupy very large footprint. So
[38:28] by using this high density XPO you can
[38:31] actually condense greatly reduce the
[38:33] footprint of the switch uh form factor.
[38:38] And here we actually uh doing our first
[38:41] uh live demo of 12.8T at uh SPO and
[38:44] welcome to stop by Terra booth to see it
[38:48] and basically we are now uh showing all
[38:51] 64 channels live and running and running
[38:54] a very low uh BR floor. So this is give
[38:58] you a feasibility demonstration. Uh this
[39:00] kind of very high channel cam
[39:02] pluggability actually can be achieved uh
[39:05] with that large investment in the past
[39:08] on the pluggable silicon photonics.
[39:12] And here's the open CPX MSA. Uh a few of
[39:14] us industry just announced uh including
[39:18] industry leaders uh on the module making
[39:20] side but also very important the
[39:22] connector community join uh this uh MSA
[39:27] and the key driver for this MSA is
[39:30] bringing the plugability to the MO CPO
[39:34] use cases. So the end customer can enjoy
[39:38] interoperability
[39:39] and also enjoy the same supply chain
[39:43] resiliency and supply security with open
[39:47] ecosystem.
[39:49] And here just to give you example how
[39:51] these open CPX mo can be used. Uh you
[39:56] can certainly this is an example you can
[39:58] actually populate these uh
[40:02] socketed engines either on the
[40:04] motherboard it become a near package
[40:07] optics or if the substrate can be opened
[40:10] up you basically can use the same engine
[40:13] on a CPO uh form factor and there are
[40:18] newer types where you can actually have
[40:21] this extension of interposer we call it
[40:24] integrated optical attachment ment. So
[40:26] you can actually have variety of options
[40:28] to attach these optical engines with a
[40:31] socket. And again uh terhop had a demo
[40:35] uh live demo in our booth on a 6.4t
[40:38] engine uh with a multisource connector.
[40:42] So welcome to stop by our booth as well.
[40:45] Uh one thing have not been talked about
[40:48] but uh we also working on is a multiore
[40:50] fiber. uh this is going to be
[40:52] anticipating a large and large number of
[40:55] fiber going to be deployed and some of
[40:57] the data center buildings networking
[40:59] building particular will have 10 or 20
[41:02] million piece of fiber installed. So
[41:04] this is the P pain point I think our
[41:06] industry have to address
[41:09] and a few of the industry leaders AFL
[41:11] Corning Sumitomo and Terrahub we now
[41:13] forming AMSA again just announced right
[41:17] before C and this is the intention is to
[41:20] foster a open ecosystem and prepare for
[41:24] future multicore fiber deployment in the
[41:26] next uh few years and we already
[41:29] demonstrated uh actually in OFC we're
[41:32] going to be join doing a joint demo
[41:33] recording uh with 1.60
[41:36] MCM native transceivers uh to start that
[41:40] uh industrywide collaborations.
[41:44] Uh here my concluding chart really
[41:46] putting everything we've done before but
[41:49] also anticipated new optics coming to
[41:52] the scene. So uh on the top line
[41:55] basically we are using the logarithmic
[41:58] scale here. You can see every uh 2x
[42:01] jumping is representing uh in a vertical
[42:03] domain but this is showing that each
[42:06] module capacity wise you you can see
[42:10] there are actually three increments
[42:11] almost simultaneously in the next couple
[42:14] years. So this is uh speed up on our
[42:16] industry to come up with the right
[42:18] solution with open community to really
[42:22] uh drive this uh uh fast speed adoption
[42:25] and innovation and large scale
[42:28] deployment and I might say uh we haven't
[42:31] really showed the upgrade uh of XPO but
[42:34] certainly we anticipate another jump if
[42:36] we go to 400 g per lane. So I'll call it
[42:39] XPO plus for now uh as a precursor later
[42:42] on we're going to work on this as well.
[42:45] uh including OpenCPX we're going to
[42:47] start with 200 gig plane but I think 400
[42:49] gig plane is not that far behind
[42:52] and then I mentioned the MCMSA
[42:55] uh I think this is again prepared for
[42:57] the next couple years I think the newer
[42:59] type of fiber will need to be deployed
[43:02] as well thank you very much
[43:07] so I mean it it's we we've got I think
[43:10] what was really interesting on this
[43:12] panel is that we have systems folks, we
[43:14] have components folks, we have
[43:15] integrators. I mean this is this is it.
[43:17] This is the economic center of scaleout
[43:20] networks and and there is a lot going on
[43:23] and a lot lot of information to pass
[43:25] through. So I have I mean starting from
[43:28] CPO it's all pluggables today. Yes, CPO
[43:31] is coming very very fast. Jensen
[43:32] speaking as we as we are sitting here
[43:35] and I was looking for the tweets to come
[43:36] out. Haven't seen anything yet but maybe
[43:39] we can uh start with that. I let me ask
[43:42] my question anyway since I've got the
[43:44] got the privilege of being the moderator
[43:47] here. So Whipple maybe we just go it's a
[43:50] poll you know everybody likes to do
[43:52] polls these days it looks like. So I
[43:54] want to do a poll say what percent of
[43:56] let's pick a year 2028 or 2029 let's
[44:00] pick 2028 what percent of 2028
[44:03] uh is going to be CPO and I'm going to
[44:06] write it down
[44:07] >> in scale out
[44:08] >> in scale out. Yeah, we're only talking
[44:10] about scale out here, right?
[44:12] >> 20%.
[44:14] >> 20%. Okay, Rada.
[44:18] >> Um I was going for a more conservative
[44:19] number. Um okay, I I let's assume scale
[44:25] out is is an artifact. Uh let's assume
[44:27] CPO in general deployment that that
[44:29] would be that would be better. Um I at
[44:32] 2028 um all of 26 27 end of 28 will be
[44:36] actually three years. Uh I think in in
[44:39] in a variety of combinations I think CPU
[44:43] uh it's not going to cannibalize the
[44:46] pluggable market. Let's let's be careful
[44:48] about that. But I think it'll add on to
[44:51] the pluggable market or the market share
[44:53] by you know uh I would say between 10
[44:56] and 15%.
[44:58] >> Okay. I wrote that down near.
[45:01] >> Okay. Now the real one 30
[45:04] >> the real answer.
[45:05] >> 30%. Wow.
[45:08] >> Wow.
[45:08] >> I I would say Jensen Huang is the
[45:10] biggest factor here. So, it depends on
[45:12] how hard he push and
[45:14] >> actually getting it to the market. So,
[45:16] >> I would say it's a big swing. It could
[45:18] be five to 20%.
[45:20] >> Okay.
[45:21] >> Depends on how successful Jensen any
[45:23] credit
[45:25] >> that's a bigger range. Maybe
[45:26] >> you guys are not giving Jensen any
[45:27] credit
[45:28] >> number.
[45:31] Okay. So, why don't I uh I I do like to
[45:35] open it out to the audience. I have a
[45:37] few other questions, but we we'll start
[45:39] with the audience first. Thank you,
[45:43] >> Ryan. Um you mentioned the need for an
[45:47] MSA for multicore fiber. Uh do you think
[45:49] we need one for holocore fiber as well?
[45:53] >> Uh I will hesitate to comment. I think
[45:56] right now the cost is so high at least
[45:58] as far as I know. So I do see challenges
[46:01] there um how to actually become more how
[46:05] to say um general utilization of power
[46:10] fiber and mostly right now it's very
[46:12] limited to DCI and scale across where
[46:16] the connection itself is so high cost
[46:19] probably they can afford to pay for it.
[46:21] Um so on my side I'll be waiting for the
[46:25] cost to go come down at least order
[46:26] magnitude.
[46:28] >> Okay. Thank you.
[46:31] >> Uh yes, hello. My name is Zian Malarios.
[46:33] I'm with QTI and there was a lot of
[46:36] discussion about obviously the growth of
[46:38] data and so forth and all these links.
[46:42] What are your thoughts around how do you
[46:44] secure this data in terms of quantum
[46:47] safety or QD? Is that part of uh the
[46:51] discussions that you're having right
[46:53] now?
[46:55] >> Who would like to take it?
[46:56] >> Um okay. Um I I'll I'll give give it a
[47:00] shot. Uh so the the the encryption part,
[47:04] right? I mean quantum otherwise uh has
[47:07] traditionally fallen on the uh the line
[47:09] line card side. Uh the the system level
[47:12] system level encryption uh but more and
[47:15] more uh they're pushing it down to the
[47:17] pluggable levels. Even DSP these days uh
[47:20] you'll see the next generation DSPs uh
[47:22] with MaxAC uh built into them. So yes um
[47:28] you know uh not not taking sides on
[47:30] algorithms
[47:31] uh picking on industry standards for
[47:34] today it'll be uh pushed down all the
[47:37] way uh into the pluggables uh that's
[47:39] happening as we speak uh a lot of our
[47:41] DSPs are being designed uh with
[47:43] encryption built into them.
[47:46] >> Yeah. Yeah please. So just to address
[47:50] one small piece of it, the generation of
[47:52] a random number. There's one thing
[47:54] lasers do and that's spontaneous
[47:56] emission and it's
[48:00] really random. I think if that can be
[48:02] leveraged that would be a good asset.
[48:04] The challenge is h how you capture um
[48:08] changes it and reduces the randomness.
[48:11] So it's a good problem to have.
[48:15] >> Okay. Thank you. Another question.
[48:18] >> Yeah.
[48:20] >> Yeah. It's very hard to see here.
[48:21] >> Yeah. Frank Chong from Southronics.
[48:24] >> Yeah. I have a quick question regarding
[48:25] the comments you guys did for the CPU
[48:27] deployment. I know we talk about CPO for
[48:30] last several years, right? Every year we
[48:31] say N plus2, right? Look like we are
[48:34] definitely
[48:36] with that. So for the for the XPO uh
[48:39] impact. So this year right we talk about
[48:42] future for for pluggable optics. Can you
[48:45] comment what is impact you think that
[48:47] will be regarding the CPO uh deployment?
[48:50] >> Yeah, thanks Frank. That's a I I'll just
[48:53] make a couple of comments, right? I mean
[48:55] it's like the one thing that is always
[48:56] obvious is new technology is going to
[48:58] replace old technology. The timing is
[49:00] what's not is not being you know that's
[49:02] the hard part. But with that said uh who
[49:05] would like to take this one?
[49:06] >> I think all of us.
[49:07] >> Yeah, maybe we do it in a in a se let's
[49:09] start with Ryan first. Yeah, please
[49:10] Ryan.
[49:10] >> Yeah.
[49:11] >> Okay. Um, if I would recommend to the
[49:14] end customer, I would say
[49:17] XPO when you can and CPO when you must.
[49:22] >> Okay. Hey, that that's a sound bite.
[49:25] >> What about Y?
[49:26] >> Uh, I think uh it's going to be
[49:29] dependent on where people are going with
[49:32] uh the CPO for the scaleup system. I
[49:35] think the volume in the scaleup system
[49:37] may may uh may change the equation but
[49:40] otherwise what Ryan said is right.
[49:44] >> U the the one of the things about this
[49:46] XPO is the interoperability between CPO
[49:49] and CPC
[49:51] uh that you you can use the same sort of
[49:53] the form factor. So that would that
[49:56] would give uh that would give people
[49:58] deploying it like okay if one doesn't
[50:00] work I'll deploy that uh uh and most
[50:04] likely we we'll start with copper always
[50:06] does and then uh deploy uh CPO as needed
[50:10] and I agree with the uh the sound bite
[50:12] uh but that that form factor allows for
[50:16] u sort of interchangeability as needed
[50:20] >> I'll address a subset Frank you started
[50:22] your question by saying we've been
[50:24] talking about it for a a long time the
[50:25] CPO um so and then you ended with your
[50:29] XPO question so let me get that out
[50:31] first yeah I think XPO will do very well
[50:33] and all good
[50:36] the one point we don't often talk about
[50:39] enough is
[50:42] the credit Nvidia deserves for grabbing
[50:45] the bull by the horn and solving the CPO
[50:48] problem from technology to supply chain
[50:52] to OSAT to advanced packaging and it is
[50:56] precisely because they spent this money,
[50:58] resources and technology to solve it
[51:01] that an open CPO is now more probable
[51:04] and more likely. So it's an interesting
[51:07] dynamic. Closed system, one big buyer
[51:10] steps in, solves all the problems,
[51:12] encourages the rest of the market. Now
[51:15] we can go open CPO.
[51:17] >> So we've seen two MSAs
[51:20] related to CPO technology.
[51:23] One is the OCI and the other one is the
[51:26] CPX open CPX. Do you think they are
[51:30] complimentary to each other or are they
[51:32] competing with each other?
[51:33] >> It's a great question. Maybe we should
[51:35] just do the
[51:37] >> sequence again. We start this way,
[51:39] please.
[51:40] >> First of all, uh, do you have any idea
[51:42] how intimidating it is to have these
[51:43] flashlights?
[51:44] >> That's very right.
[51:46] >> And then a question coming from a
[51:47] physicist I admire.
[51:50] Um, no seriously I I see nothing but
[51:53] complimentary nature because the OCI is
[51:55] saying take four lambda, use NRZ, use
[51:59] birectional optics. Um, and what OpenCPX
[52:03] is saying, whatever your optics is, make
[52:05] sure it's on a socket, make sure it's
[52:07] friendly to copper. So, the two go hand
[52:09] in hand and help each other.
[52:12] >> Um, yeah, I I agree with Ripple. I
[52:15] think, uh, by design, the um, OCI does
[52:18] not have a form factor. It's it's got
[52:21] specifications for interoperability. So,
[52:23] they have to sort of coexist in some
[52:25] plane.
[52:26] >> Uh, I'm going to disagree. I think
[52:29] they're directly competitive uh with
[52:31] each other because one is relying on a
[52:33] high-speed serial link to reach the
[52:35] optics and the other one does not.
[52:38] >> Yeah, thank you
[52:39] >> Brian. So this is actually come back to
[52:42] my earlier comment is um new technology
[52:44] probably needs some proving ground and I
[52:47] think there's certainly a slow and wide
[52:50] promise uh down the road can be a very
[52:52] low power options and I think that's
[52:55] just physics right uh you can't argue
[52:57] with physics
[52:59] >> uh but to get a deploy I think we need
[53:01] to find the right way to deploy it so
[53:03] actually I I I have a bullet there say
[53:06] even XPO can have the slow and wider
[53:08] optics embedded in there because that's
[53:10] how maybe the first couple generation
[53:13] you can get OSP you can get SPO to embed
[53:16] these technology and start to prove it
[53:18] out and you may get incremental halfway
[53:21] saving but you not get 100% saving but
[53:23] at least that's a way you can get it
[53:24] deployed and even on the server side
[53:27] versus the switch side you also can have
[53:29] one side of deployment one form factor
[53:32] the other side deploy another form
[53:34] factor so there's lots of way we can
[53:36] actually mix mismatch the different
[53:38] things uh coming back to the two MSA I
[53:41] think those two MSA CPX can naturally
[53:44] also incorporate a slow and wider uh
[53:47] optics as well so I don't see them as
[53:50] competing so potentially can be
[53:52] complimementaryary as well
[53:53] >> yeah I I would say what Ryan says right
[53:57] uh you can use OCI and kind of non in
[54:00] using still high-speed serial links but
[54:03] the full benefit of using slow and wide
[54:05] is getting rid of high-speed serial
[54:07] links
[54:07] >> that's great yeah thank
[54:09] There's 10 minutes. Yeah. Um,
[54:11] >> hey Rajie Control Broadcom Vipple. I'm
[54:14] confused. When you talked about CPO, you
[54:16] only talked about Nvidia solving all the
[54:19] problems.
[54:21] >> Thank you, Rajie.
[54:23] >> Yeah, I didn't like that comment either.
[54:27] >> Highly offensive.
[54:30] >> May I
[54:31] >> Moving on? In my
[54:34] in in my defense,
[54:37] these two are the members of the N plus2
[54:39] club.
[54:41] CPO is two years away. CPO is two years
[54:43] away, right? So I call, you know, we've
[54:45] been doing that for n number of years.
[54:47] Uh it was truly um outstanding effort by
[54:51] Nvidia that that's that went from a to
[54:57] >> we going to boot him off the stage soon,
[54:58] guys. Watch. Okay, let's change the
[55:01] topic.
[55:02] >> Another controversial one maybe uh the
[55:05] rather you're the only one who spoke
[55:07] about coherent light and scale out. So
[55:09] why don't we start with you and we'll
[55:11] get comments from everybody else.
[55:13] >> So um when we put the slide together I
[55:15] thought um I'll probably have to come on
[55:17] stage and defend coherent light. But
[55:19] after I heard the first panel beginning
[55:21] of the beginning of the day and like wow
[55:24] all these people want coherent light why
[55:26] didn't I add another slide to it? Uh
[55:29] hey, no laughing. Uh so
[55:34] it's it's uh
[55:37] it's it's always a case of is it going
[55:38] to be higher power, right? It's it's
[55:41] whenever you say coherent uh people get
[55:44] worried. Whenever you say CPO, people
[55:46] get worried about reliability. When you
[55:48] say coherent, they're worried about
[55:49] power. No. Uh engineering will get
[55:52] around. And as I said when I left the
[55:54] stage, it's not it's not if it's going
[55:56] to be deployed. there are uh deployments
[55:58] as of now and as part of a larger uh
[56:02] scale across strategy scale out networks
[56:05] would change and um it's it's our firm
[56:09] believe that um at 400 is where the real
[56:12] battle is uh just just because uh the
[56:16] physics is not in your favor uh but you
[56:18] know IMD will find ways around it right
[56:21] the incumbency is is is really strong um
[56:25] wouldn't say too strong it's really
[56:26] strong uh but uh there is an opportunity
[56:30] for it to coexist first uh and I think
[56:33] um I think it'll happen it'll happen in
[56:35] a larger scale than is happening today.
[56:37] >> Not to put you on a spot but cost and
[56:38] power where do you think the
[56:40] state-of-the-art is compared to IMD like
[56:44] to go around that that topic as well.
[56:47] >> All right. Uh defending coherent wasn't
[56:50] wasn't in my cards today but it's it's
[56:51] it's a great business to be in. Um okay
[56:55] there there's one other advantage to to
[56:58] a coherent lambda
[57:02] you take an F fr FR4 it's a single
[57:04] lambda in coherent right that there is
[57:06] that advantage where where you you there
[57:09] are things that you save in in chip
[57:12] design etc because people look at
[57:14] coherent and say oh god it's it's so
[57:16] much more power but then it's 4x the
[57:18] density uh of PAM 4. So there's that
[57:21] working for it. As we get to a a point
[57:24] where where there only that many lambdas
[57:27] that you want to work with, uh coherent
[57:30] would would become more and more and
[57:33] more attractive. But I think power per
[57:35] lambda wise if we if we considering the
[57:39] whole module if we were to use standard
[57:43] DFBs and if we would use standard
[57:45] components and I do not see it being
[57:49] um it won't be the same I do not see it
[57:51] being considerably larger if you pick
[57:54] the right right coherent uh DSP modes.
[57:58] >> Yeah Ryan please.
[57:59] >> Yeah. So I'd like to comment a little
[58:00] bit defending coherent light. Uh I
[58:02] actually did not explicitly talk about
[58:04] it but it's on my slide actually showing
[58:06] coherent need to be a strong play. Uh
[58:09] particularly if you move to 400 gane uh
[58:12] simple physics again I mean if you use
[58:14] uh FR4 type uh you probably barely go
[58:17] 500 meters. Now you have choices there.
[58:20] How do you extend that distance? uh
[58:23] they're not just one way but I think
[58:25] coherent light is a pretty
[58:26] straightforward to compensate for
[58:28] dispersion to give you more link budget
[58:31] and that also come into play if you
[58:33] start to have OCS in part of that play
[58:36] then that extra 34 DB going to be very
[58:39] important so again another I guess catch
[58:44] word is deploy IMD you can
[58:47] >> but you're like you must right so I
[58:51] think we are running that us pretty
[58:53] soon.
[58:54] >> Something about consistency here. Yeah.
[58:56] Well done, Ryan.
[58:58] >> M. Do you have any comments on this one?
[59:00] >> Um, no. I um the only thing I'd say is
[59:02] yeah, coherence going to come from the
[59:04] long uh longer distances, continue to
[59:07] shrink the distances, but I think some
[59:09] of the slower slower wide technologies
[59:12] will come from the very very short
[59:13] distances like hopper level distances
[59:15] and come up and there will be some
[59:16] meeting point in the middle that will
[59:18] squeeze regular IMD
[59:20] >> ripple. No, but I do have a new topic
[59:22] for discussions.
[59:24] >> Hopefully, it's a controversial one.
[59:26] Yes.
[59:27] >> Oh. Um, near
[59:29] >> Yeah.
[59:30] >> teach me how to calculate how many
[59:32] transceivers per GPU given the highix
[59:35] network design.
[59:37] >> I get this question all the time and I
[59:39] keep saying it depends and it doesn't
[59:40] work.
[59:41] >> Ah, yeah. You've got to do your radics
[59:43] of your switch, right? And then you
[59:45] divide that by two. That gives you your
[59:47] first level switch so that you can
[59:50] connect a certain number of first level
[59:51] switch then uh if you run if you need to
[59:54] connect more you need a next level next
[59:56] level that's assuming that each GPU is
[01:00:00] connected directly uh directly to the
[01:00:02] switch but I think what people are and I
[01:00:04] alluded to in the talk is uh the ability
[01:00:07] to use multi-rail architecture where the
[01:00:10] clusters of the GPUs can talk among
[01:00:13] themselves to exit to the correct GPU
[01:00:16] And there therefore I only have to
[01:00:18] connect to one level switch.
[01:00:20] >> So is it three or five or what?
[01:00:22] >> Uh so typically uh typically uh right
[01:00:25] now it's uh six is six is the upper end.
[01:00:29] Four is four is typical and it could in
[01:00:32] the future go get all the way down to
[01:00:33] two
[01:00:35] >> if it's nonblocking. Isn't it just
[01:00:36] number of tiers of switch right? You can
[01:00:38] almost
[01:00:39] >> Yeah. roughly.
[01:00:40] >> Yeah. Yeah. Yeah. If it's a single tier
[01:00:42] then it's just two. One at the GPU one
[01:00:44] at the switch. depends on the
[01:00:45] architecture.
[01:00:47] >> Yeah, it doesn't matter how you connect
[01:00:48] but it's just as long as non-blocking
[01:00:50] it's just number of tier of switches
[01:00:52] >> and the bandwidth uh ratio between scale
[01:00:55] up and out will remain roughly 9 to1 or
[01:00:57] 10 to1
[01:00:58] >> uh roughly where people seeing 10 to one
[01:01:01] but you know does that change a little
[01:01:03] bit plus or minus a little bit possibly
[01:01:05] but it seems like 10 to one is a good
[01:01:07] good number to use.
[01:01:08] >> That's great. Thank you. Let's move on
[01:01:10] to the audience because we're having too
[01:01:11] much fun here talking ourselves.
[01:01:17] >> We're actually getting very close to
[01:01:18] lunch time. So you have a really really
[01:01:20] cool question to keep them exciting.
[01:01:23] >> So what is your readiness in terms of
[01:01:27] looking at fulfilling the demand? The
[01:01:29] numbers we're seeing as an industry is
[01:01:32] just, you know, like nothing we've ever
[01:01:34] seen before. So maybe we'll start with
[01:01:36] Ryan this time and work our way this
[01:01:38] way.
[01:01:40] Um
[01:01:41] right now from our point of view cap raw
[01:01:45] capacity itself on the final
[01:01:49] module product level is not a
[01:01:50] limitation. I think right now the whole
[01:01:53] industry do have constraints on various
[01:01:56] choke point is on the raw component
[01:01:58] side. I think traditionally our
[01:02:00] particularly 3FI
[01:02:03] fabs is certainly not at the scale of
[01:02:06] Camos and I hope that's going to change
[01:02:10] um you know from these really high
[01:02:12] demand but I think that's the constraint
[01:02:14] where right now industry is experienced
[01:02:16] is how much capacity you can put in to
[01:02:18] scale it up very quickly and I think
[01:02:21] that's really unique challenge in in in
[01:02:23] the last couple years because we never
[01:02:25] see this kind of ramp up like new
[01:02:28] generations And just you know basically
[01:02:30] from zero to millions
[01:02:32] literally in one year and this is just I
[01:02:35] I think the challenge we face and we
[01:02:37] work with all the partners right uh like
[01:02:39] component level including you know
[01:02:42] actually everybody on on this panel here
[01:02:45] so we all collaborate try to run that up
[01:02:47] um hopefully next year is better but uh
[01:02:49] this year is very constraining right now
[01:02:52] >> um yeah uh indium phosphide lasers is a
[01:02:55] major problem 100% uh we are trying to
[01:02:58] address it. We're scaling but everybody
[01:03:00] all our competitors are trying to
[01:03:02] scaling but it's it's a it's a challenge
[01:03:04] to get uh to the right scale in the
[01:03:06] right time.
[01:03:08] Um yeah I I it's it's the material
[01:03:11] constraint. I think the uh the module
[01:03:14] level assembly services um could could
[01:03:17] use some scaling but it's mostly the
[01:03:19] materials doesn't matter. You know you
[01:03:21] could talk about a a capacitor being
[01:03:23] missing is still missing. So it's it's
[01:03:25] right across industry because the
[01:03:27] optical
[01:03:28] modules tend to use somewhat of a higher
[01:03:30] performance components generally because
[01:03:32] of the data rates etc. So they they are
[01:03:34] at a premium right now.
[01:03:36] >> Whipple. Yes, we are constrained doing
[01:03:39] the best.
[01:03:41] >> A simple answer.
[01:03:44] Okay. All right. With that, I got to
[01:03:46] thank our panelists. Thanks a lot for
[01:03:48] your presentations and uh
[01:03:52] >> um and the interesting discussion. Thank
[01:03:54] you very much. Thanks. Thanks a lot.
[01:04:09] Everybody talking about scaling AI,
[01:04:14] but the data center's choking deep
[01:04:17] within.
[01:04:21] Copper running hot. Yeah, the signal's
[01:04:24] getting thin.
[01:04:26] So we flip the switch now.
[01:04:30] Optics is in
[01:04:33] bandwidth climbing fast. Racks are
[01:04:37] running red.
[01:04:39] Cloud demand exploding overhead.
[01:04:45] Pluggables fading as the limits closing.
[01:04:51] Cold package light is how we win.
[01:04:56] It's photo
[01:04:58] baby 2026
[01:05:02] riding that lightweight
[01:05:05] doing new tricks
[01:05:08] from the fiber in the ground to the chip
[01:05:10] in my hand. We make that sunshine jump
[01:05:13] on command. Yeah. Bonx baby
[01:05:20] 2020.