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

https://www.youtube.com/watch?v=PYKCeXl919o

[00:00] Highlights from the Optica Executive
[00:02] Forum at OFC. Thanks to these sponsors.
[00:20] We're releasing seven sessions from
[00:22] Optica's Executive Forum because the
[00:25] world is watching. Feedback encouraged.
[00:29] Now, enjoy.
[00:31] This next panel is about a scale up
[00:34] scale up of photonic technologies. All
[00:37] of you are checking the phone and see
[00:39] what your friend Jensen is talking
[00:41] about. Chris Fisher is going to be
[00:43] moderating this event, this next panel,
[00:46] and see whether this impact of the GTC
[00:50] has an effect here on the photonics
[00:52] ecosystem. Chris Fisher, thank you very
[00:55] much for moderating this panel. The
[00:56] floor and the attention of everyone is
[00:58] yours. Jose, thank you for the kind
[01:00] introduction. Good afternoon and welcome
[01:02] to the scale up panel. We had scale
[01:05] across, we had scale out.
[01:07] Now, the toughest one because, as you
[01:09] might have noticed in scale out, scale
[01:11] across, optics is the default
[01:14] internetworking technology. At scale up,
[01:16] it's still copper,
[01:18] but we're getting air cover for those
[01:20] who keep track. Uh Jensen Huang
[01:23] announced today that Nvidia is going to
[01:25] push CPO into scale up with the Vera
[01:29] Rubin ultra front generation, so the NBL
[01:32] 576. So, it's coming.
[01:35] Um we have another all-star lineup here.
[01:38] Most people know most of the speakers,
[01:40] but I'll introduce them briefly anyway,
[01:42] and then we'll do the same thing. We'll
[01:44] We'll have them come up one by one.
[01:46] Hold your questions for the panel, then
[01:49] we'll go up and we'll all sit down and
[01:51] take the panel uh discussion. So, first
[01:54] up, uh Dritan Al Dweina is going to
[01:56] start start us out setting the tone with
[01:59] Meta as one of the end customers here.
[02:01] Drew is a principal principal engineer
[02:03] at Meta focused on integrated optics for
[02:06] AI.
[02:07] Prior to Meta, he was an early member of
[02:09] the Silicon Photonics team at Intel.
[02:12] Then we have Dave Lozovsky.
[02:15] Dave was the co-founder and CEO of
[02:17] Celestial AI, which was acquired by
[02:19] Marvell as most of you probably know. He
[02:22] is now the EVP and GM of the data center
[02:24] networking business group at Marvell.
[02:28] Following him will be Peter Winzer.
[02:30] Peter was co-founder and CTO of Nubis
[02:33] Communications,
[02:35] which was acquired by Ciena,
[02:37] and he's now the VP system architecture
[02:39] at Ciena.
[02:41] And then last but not least to round us
[02:43] off will be Andy Bechtolsheim.
[02:46] Again, most of you know him and he's a
[02:48] co-founder and chief architect at
[02:50] Arista. Prior, he was one of the
[02:52] co-founders of Sun Microsystems and also
[02:55] founder of Granite Systems, which was
[02:57] acquired by Cisco a while ago.
[02:59] With this, I'll welcome Drew to the
[03:01] stage.
[03:04] Can you hear me? Yes?
[03:06] Perfect.
[03:07] All right, thank you very much for being
[03:09] here. I feel a little bit like there
[03:11] should be a target on my chest after
[03:13] that last discussion where Vlad was
[03:15] instigating with Hong and and now it's
[03:17] sort of set up somehow this discussion
[03:19] or or battle there. I hope that um
[03:22] you'll find that I'm in greater
[03:23] agreement with Hong than than
[03:25] disagreement. Um
[03:29] but we'll see. Challenge me at the end,
[03:30] I guess.
[03:31] Uh I I
[03:33] Generally, people put up co-authors on
[03:36] slides and you don't think so much about
[03:38] it, but I'd like to call out
[03:39] specifically the uh co-authors on
[03:41] on these slides in this talk, Siamak
[03:43] especially, and I this is buddy Afshin
[03:46] who have been doing most of the CPO work
[03:48] that we'll talk about, at least their
[03:49] liability work, and Darren who leads uh
[03:51] entire effort.
[03:54] Meta is an end user. Meta isn't a laser
[03:57] fabricator. We have in in fact not a
[04:01] chat GTP
[04:03] chat chat TP GPT
[04:06] um
[04:07] product that we serve, but we do serve
[04:10] an enormous number of users on a regular
[04:11] basis. Reliability means an awful lot in
[04:14] the ecosystem of products that we deploy
[04:17] and the infrastructure upon which we
[04:19] build it. Um 200 building WhatsApp users
[04:23] or messages in a regular day. It's
[04:25] It's crazy what we do.
[04:27] AI because it's not an an agent or at
[04:30] least the service that we offer to you
[04:32] today specifically, it's built into all
[04:35] these different apps. It's built into
[04:37] the apps in which we serve you, Facebook
[04:40] feed, Instagram feed, etc.
[04:42] So the recommendation engine, but also
[04:45] you know, inherently and more even more
[04:47] and more deeply into Meta Quest. You'll
[04:49] see more and more sort of AI
[04:52] agents available to you to help you do
[04:54] things, to sort things, to deliver to
[04:55] you sorts of content that that you come
[04:58] to expect from apps from Meta apps.
[05:03] The infrastructure required to sort of
[05:05] support that number of users and the you
[05:07] know, very promising future that AI
[05:10] represents for us and I think I will
[05:12] speak more broadly for the entire
[05:14] industry
[05:15] is is staggering. I just told somebody
[05:18] and they sort of laughed at me, but it's
[05:19] true, right? I mean, we have plans today
[05:23] to deliver many gigawatts of compute on
[05:25] a regular basis.
[05:27] I think even last year we said we're
[05:28] going to have more than 1.3 million GPUs
[05:31] deployed.
[05:32] We announced CapEx this year of greater
[05:34] than $115 billion, which was greater
[05:37] than the 70 billion last year and
[05:38] greater than the 35 billion the year
[05:40] before. And we're at best what, 20% of
[05:44] the hyperscaler market? So we're looking
[05:46] at numbers which are in excess of half a
[05:48] trillion dollars worth of capital
[05:50] equipment being put into AI
[05:52] infrastructure across the industry.
[05:55] I was telling gentlemen over there, it's
[05:56] actually hard to spend this much money.
[05:58] And you know, you you should you should
[06:00] laugh at this part as a good problem to
[06:01] have.
[06:02] Um but to actually I mean I'll I'll I'll
[06:05] go back to the the CTO panel that was
[06:07] here, right? Matt and and Hong and
[06:09] and Julie talked about how difficult it
[06:11] is to ramp, how it's sort of significant
[06:13] it is to sort of be able to deliver
[06:14] these volumes at this quality, this
[06:16] consistently on such a fast development
[06:18] cycle.
[06:19] We appreciate that, right? We're the
[06:21] ones who have to actually absorb it into
[06:23] the fleet and build the capacity on top
[06:24] of that, right? We're we're believe it
[06:26] or not, we're in this together. So,
[06:27] trying to give you the visibility is
[06:29] part of what we're here to do today.
[06:31] Uh but but you know, believe it or not,
[06:33] even Meta doesn't or Mark doesn't have a
[06:34] crystal ball. So, um so, challenge as
[06:37] you want, but don't expect perfect
[06:38] answers.
[06:40] I will say uh I continue to be stunned
[06:42] by uh the Hyperion data center that we
[06:44] announced in Louisiana,
[06:46] uh ability to scale up to 5 gigawatts
[06:49] over its lifetime.
[06:50] Uh as Mark said, it covers a significant
[06:52] part of the footprint of Manhattan.
[06:55] Um we which is, you know, we're not
[06:58] we're not building buildings anymore.
[06:59] We're building sort of urban centers
[07:00] that have uh AI computers in them. So,
[07:02] maybe Jensen's right, right? These are
[07:04] AI factories at this point that we build
[07:06] upon it.
[07:07] Uh it's not the only one, right? We've
[07:08] already announced plans in
[07:10] Hyperion and and the Prometheus, right?
[07:12] Ohio, Texas, Indiana were all sort of
[07:14] announcing this is part of the growth
[07:16] curve uh of which Meta is a part.
[07:19] Um to do that, we partner with you, the
[07:22] broader industry, right? I mean, you can
[07:23] go and see our latest announcement about
[07:25] our own internal silicon for GPUs at
[07:28] ai.meta.com. I encourage you to do that.
[07:30] That's just sort of came out.
[07:32] Um but that capacity doesn't just get
[07:34] built on one
[07:35] company's capacity, right? We have um
[07:38] partnership agreements with Nvidia and
[07:39] AMD and recently announced with Corning
[07:41] as well to secure the types of capacity
[07:43] that are required to deliver this sort
[07:45] of infrastructure across all our use
[07:47] cases.
[07:51] This talk is supposed to be about scale
[07:53] up and I promise I'll get to it, but
[07:55] the you know,
[07:57] largely as we build out these these
[07:59] grand, you know, city-size AI data
[08:02] centers, there's an awful lot of
[08:04] networking which goes into that to build
[08:05] upon the GPU and compute capacity that's
[08:08] in there.
[08:09] Um it certainly is putting a great deal
[08:11] of
[08:12] pressure onto the network, uh both
[08:14] capacity to sort of scale as required
[08:17] for sort of cost appropriate ways, but
[08:19] also the power.
[08:20] Um LPO and CPO and generally integrated
[08:23] optics and as we'll talk about OCI
[08:25] uh seem to be very interesting ways to
[08:27] save power. I will agree largely with
[08:30] Hong and I'm going to wave my hands cuz
[08:32] I can't see anything. I'm assuming she's
[08:34] here, but but you will believe it's it's
[08:36] the grand edifice of of Google and and
[08:38] Hong's insight.
[08:40] Um network is not the major contributor
[08:42] of power and that that's a really good
[08:44] thing, right? Because really what we
[08:45] want is a GPU, the flops.
[08:47] Um
[08:48] nevertheless, it is a significant part
[08:50] and the more you reduce the power,
[08:52] uh the better off you are.
[08:54] If 5% of 5 gigawatts is still an awful
[08:56] lot of GPUs and that's sort of the way
[08:58] we think about it, right? The law of big
[08:59] numbers still gives you those sort of uh
[09:02] interesting outcomes.
[09:05] I think that we have and are answering
[09:07] this significant question that we have,
[09:09] which is, you know, can we build these
[09:12] more advanced non-re-timed optics and
[09:14] save ourselves power just because we're
[09:16] moving electrical bits around a sort of
[09:17] short channel. I think we've seen that
[09:20] over time. LPO was what, three
[09:23] three OFCs ago? That sort of got
[09:25] announced, maybe four.
[09:27] CPO, I guess became
[09:30] popular two
[09:32] uh C OFCs apart.
[09:34] Uh I think the answer or the question
[09:35] that we have to answer for ourselves and
[09:37] with your help is should we be building
[09:39] these things? Should we be deploying
[09:41] these things? In some sense is the juice
[09:42] worth the squeeze? Because they all come
[09:44] at a cost. You lose something by doing
[09:46] this, whether it be
[09:49] uh
[09:50] availability or telemetry or
[09:52] serviceability or something. And those
[09:54] those costs are real. And when we talk
[09:56] about TCO and I I get yelled at for
[09:58] this, those costs get
[10:01] um
[10:02] rolled up into the way that we consider
[10:04] what the impact on our capital equipment
[10:06] and our deployment plans is.
[10:08] So, from our perspective, the biggest
[10:11] unanswered questions really are and I
[10:13] don't want to call this out
[10:13] specifically, the RAS questions. So,
[10:15] reliability, do these things work? Do
[10:18] they work consistently?
[10:20] Um do they break and when they break,
[10:22] how do they break? Availability is the
[10:24] other part of that, right? I mean, how
[10:26] frequently do they break? How do they
[10:28] break? Are they serviceable or are they
[10:30] not? Do they take away an enormous
[10:32] amount of your capacity at like with one
[10:34] little failure point? Or they more like
[10:36] pluggables where you get, you know, four
[10:38] channels as opposed to 400 or or or
[10:40] 4,000?
[10:42] Serviceability is the other part of the
[10:44] RAS equation, right? How do you fix it?
[10:46] How long does it take? I think Hong
[10:48] Again, I'm I'm using Hong as as as the
[10:50] great sort of voice of reason about what
[10:53] data center operators. If it takes you a
[10:55] week to service a switch and you lose X
[10:58] amount of capacity because it takes you
[11:00] that week to do that, then you need X
[11:01] amount of those that excess capacity,
[11:04] right? A switch takes out 10 racks, then
[11:06] I have to have 10 racks in spare for
[11:08] every switch that's down. This is the
[11:10] sort of capacity and why we think about
[11:11] TCO when it comes to RAS.
[11:17] Generally, I think as the optics
[11:19] industry, we think about pluggable
[11:20] transceivers, we think about lasers and
[11:22] die and like fiber attach and maybe a
[11:24] wafer, which is really cool. This is the
[11:27] product. This is the product that NetApp
[11:28] buys and installs in their
[11:30] infrastructure. This is what we need to
[11:32] make real and then keep running as a
[11:35] reliable product in the data center that
[11:37] we can drive our capacity on. If you
[11:40] look at this one
[11:42] I wish I had a laser pointer.
[11:44] You'll see in the center there's two
[11:46] racks, a compute rack and basically a
[11:48] network rack. And then there's three
[11:50] racks on either side of it, which are
[11:53] the cooling racks. So so for us, one
[11:55] GP300
[11:57] rack takes five rack positions to deploy
[12:00] and service.
[12:02] So this represents power, networking,
[12:04] and cooling on top of the compute.
[12:07] So when we think about sort of
[12:10] what are the services that are required
[12:11] from a network perspective to support
[12:13] that infrastructure? Well, I mean
[12:14] there's obviously front out front end
[12:16] and scale out links, hundreds of meters,
[12:18] rack to rack across a data center in
[12:20] this sort of enormous factory.
[12:23] Um
[12:24] but there's also this sort of scale out
[12:26] domain cabling that that comes from the
[12:28] network rack. And then what we're going
[12:30] to talk about today is the scale up
[12:32] domain cabling.
[12:33] Currently, at least in the GP300 rack
[12:36] infrastructure, is a high-speed
[12:38] electrical backplane. Again, it's it's
[12:41] copper, right? It's hard to imagine a
[12:43] more mature industry than extractive
[12:46] sort of copper from the ground and
[12:49] pulling it into wires. We've been doing
[12:50] that for literally millennium.
[12:54] Um it it's well understood. It's going
[12:56] to stay there forever.
[12:58] 400 gig is coming. We all sort of
[13:00] understand that. If the challenge is
[13:01] replacing electrical backplanes with
[13:03] optical backplanes, which is which I'm
[13:05] positing is, then that's the gold
[13:07] standard at which we need to judge
[13:09] ourselves.
[13:12] So why would we consider that? I mean
[13:13] this is sort of the question we ask
[13:14] ourselves, right? We have this sort of
[13:16] known good copper infrastructure, et
[13:18] cetera.
[13:19] Um the challenge becomes large scale up
[13:22] domain sizes.
[13:23] Um
[13:24] Jensen I think often talks about 72
[13:27] nodes, scale-up domains, single
[13:29] uh single-wide racks that uh support 72
[13:32] accelerators and they're the
[13:34] standard these days.
[13:36] Uh
[13:37] we have uh deployed these things. You
[13:38] can see some of the characteristics
[13:41] around that.
[13:42] Um but we're exceeding the capability of
[13:46] uh standard single-wide rack to deliver
[13:48] the types of performance across the
[13:50] scale-up domain that we want.
[13:53] At OCP last year, Meta and others uh
[13:56] proposed or or described, demonstrated
[13:59] an ORW rack, which is a double-wide
[14:01] rack. So, now what we've done is we've
[14:03] basically taken twice as many GPUs,
[14:05] we've stuck them in the same sort of
[14:07] copper radius that you can achieve with
[14:09] a sort of electrical backplane copper um
[14:11] cable.
[14:13] Uh
[14:14] and we've been able to double the
[14:15] density
[14:17] of GPU compute in a copper-based
[14:20] scale-up domain.
[14:22] The challenge, of course, is that it's
[14:24] never going to be enough. Everybody who
[14:25] draws a curve sort of draws this
[14:27] exponential sort of curve. And so, the
[14:29] question becomes what happens when you
[14:30] need more than 144? What do you do when
[14:33] you need a 256 node accelerator?
[14:35] Well, excuse me, scale-up domain.
[14:39] Uh the the problem is you start looking
[14:40] at what that could be in an electrical
[14:42] rack and you're putting a megawatt in in
[14:44] a rack. I'm reminded of this sort of
[14:46] description that was given out to the
[14:47] end of the last century called smoking
[14:49] hairy golf balls. In that what you
[14:50] Where the the the author suggested that
[14:53] the the limits of compute were smoking
[14:57] wanted to do was put as much compute as
[14:58] you possibly could as close as you
[15:00] possibly could so that you could reach
[15:01] it with an electrical link. And the
[15:03] limits of that sort of golf ball or that
[15:05] 3D torus of compute connected with
[15:07] electrical would be the point at which
[15:09] it got so hot it melted or at least
[15:11] started to smoke. And then because that
[15:13] would never be enough, you'd have the
[15:14] smoking golf ball that you need to
[15:15] connect to each other. So, the network
[15:16] of the future was smoking hairy golf
[15:18] balls where the hair was the fiber.
[15:20] So, this is you know, that's a step
[15:23] towards that. Ideally, we wouldn't do
[15:25] that. Ideally, things like optical
[15:28] compute interconnects or optical
[15:29] backplanes allow us to take these large
[15:31] constructs of 256 or more nodes in a
[15:35] scale-up domain, connect them with
[15:37] fiber, and suddenly locality in the
[15:39] scale-up domain no longer is driven by
[15:41] copper interconnects.
[15:45] I think of it as a great vision,
[15:47] but but I want to be honest and fair and
[15:51] and and again, I mean, I'm sort of using
[15:53] Hong as the boogeyman, but I
[15:54] do want to call her out as a very
[15:57] thoughtful and obviously successful
[15:59] engineer who is raising very very real
[16:01] and rational questions about the way in
[16:03] which we talk about optical scale-up as
[16:05] a competing technology for copper.
[16:08] Cable backplanes are the standard of
[16:09] excellence, period, full stop, right?
[16:12] Nobody has an optical backplane right
[16:13] now, and there's a reason for that. The
[16:15] cost and the power of copper-based
[16:17] backplanes is lower. Reliability
[16:20] generally is higher.
[16:23] However, we're constrained now, right?
[16:25] Now, everything has to go in a copper
[16:26] sort of radius and and, you know, if
[16:28] there's one thing that optics does
[16:30] better than copper is a low-loss
[16:31] channel, right? The low-loss optical
[16:34] fiber is sort of the solution to
[16:36] everything if you need to get transmit
[16:38] data over distance. The question in this
[16:40] case is, you know, do you need that
[16:42] distance?
[16:44] If we're going to do this, if we're
[16:45] going to actually going to do some
[16:46] conversion of electrical backplanes to
[16:48] optical backplanes and have optical
[16:50] scale-up that works, we need a robust
[16:51] ecosystem. It can't be one player, it
[16:53] can't be the greatest sort of modulator
[16:55] since, you know, God invented
[16:57] modulators,
[16:58] but it's available on sort of one single
[17:00] supplier. So, we need an ecosystem. We
[17:02] need interoperability if we have that
[17:04] ecosystem, right? Vendor A silicon good
[17:06] to vendor B silicon has to work, and it
[17:08] can't be a single generation. We can't
[17:10] put all the effort and and work into
[17:12] qualifying that thing, and then next
[17:14] generation we have to do do that work
[17:15] and all that effort in qualifying again.
[17:17] Um I think even Matt and the and the
[17:19] other CTOs talked about the sort of rate
[17:22] of rapid change. We're I mean I
[17:24] you know, we talked about our MTIA road
[17:26] map for silicon. We have like something
[17:28] like a 6-month cadence now for new
[17:30] silicon coming out. If we have to
[17:32] innovate new optics every 6 months then
[17:34] then, you know, you guys know this
[17:35] better than I do. We're we're not going
[17:36] to get there.
[17:39] Because RAS is so important to us, we're
[17:41] doing everything we can to sort of
[17:42] collect data on RAS so that we can both
[17:44] feed it back into the system, which
[17:46] means you, but it also into our own
[17:49] calculations for what our capacity due
[17:51] to availability really means. So, we
[17:54] have done an awful lot of work and we'll
[17:55] report some more now on the available
[17:58] CPO system at the moment, 51.2 T CPO
[18:01] Bailey systems, to try and address the
[18:03] sort of chicken and egg problem where
[18:05] you need an awful lot of things deployed
[18:06] to actually get the number of
[18:07] reliability that you want, but you can't
[18:09] deploy an awful lot of things until
[18:10] you're reliable enough to actually do
[18:11] them.
[18:12] So, we'll um
[18:15] When you hear me say CPO Bailey, what
[18:17] I'm trying to say is integrated optics,
[18:18] which means optics close to you or onto
[18:21] the package in some sort of like
[18:23] well-packaged way for which failure
[18:24] really matters a lot.
[18:26] Previously,
[18:27] uh at ECOC of last year, we reported
[18:29] something like a 5x improvement of CPO
[18:32] reliability versus pluggable module
[18:34] reliability, which was encouraging.
[18:36] However, it was, you know, 17 million
[18:39] device hours against 2 million device
[18:41] hours or so for pluggable modules, um
[18:44] which wasn't enough to give us sort of
[18:45] the level of confidence that we wanted,
[18:47] so we keep going.
[18:49] We have uh now a much larger-scale
[18:52] system, a test cluster, deployed in our
[18:54] I should say lab infrastructure. Let's
[18:58] be clear about that. This is not a
[18:59] product system data center running your
[19:01] Facebook feed. Uh we've collected over
[19:03] 50 million hours of uh equivalent 400
[19:06] gig um link data on this uh collection
[19:09] of data.
[19:10] Uh and and generally we think of this is
[19:12] a phase two Bailey system uh, operating
[19:15] under normal operating conditions, which
[19:17] is about 22° C versus the previous 40°
[19:21] C. So, the data that you'll see, and
[19:23] I'll show you from phase one systems, is
[19:25] an accelerated lifetime data where we
[19:27] would expect reliability to be lower.
[19:29] Um,
[19:30] and then there's an updated build
[19:32] standard, but largely the same amount of
[19:34] uh, of uh, the the same technology uh,
[19:38] deployed there.
[19:40] Our data has been updated. If you look
[19:42] at sort of these four lines here, we're
[19:44] trying to tell the story, I I would say.
[19:47] We're trying to draw
[19:48] uh, conclusions
[19:50] based on the available experimental data
[19:53] that we have about what long-term
[19:55] integrated optics
[19:57] uh, reliability will be based on the say
[19:59] Bailey CPO systems that we have.
[20:02] So, if we look at this, right, the the
[20:04] baseline here, like for like A for B
[20:06] comparison, is the pluggable optical
[20:08] modules that you see here, two by 400
[20:10] gig FR4 modules.
[20:13] Okay, I think I have two slides left.
[20:15] This I'm going to I'm going to keep
[20:16] talking now, so you're going to have to
[20:17] tackle me off the stage. Uh, we've got
[20:19] something like 8 million device hours at
[20:21] this time point, and an MTBF about 0.7,
[20:24] which is better than we
[20:26] had uh, shared earlier, but still
[20:29] not great.
[20:30] If we look at all the failures that we
[20:32] collected in over 40 million hours uh,
[20:34] equivalent 400 gig device hours of CPO
[20:37] phase one, it's about 1.5 or so. So,
[20:40] there's about a 2x improvement in
[20:42] reliability from CPO.
[20:44] Again, uh, we'll have Andy will come up
[20:46] later, and I think one of the great
[20:47] things that Andy said recently over the
[20:49] last year is that like a smaller bag of
[20:51] parts is more reliable than a larger bag
[20:53] of parts. So, maybe that's part of of
[20:55] this, right? I think, you know, we
[20:57] should be clear about that.
[20:59] I will call out that again, right, we're
[21:00] trying to extract a signal here based on
[21:02] the experiment that we have. If we
[21:04] exclude this known ELSFP issue due to
[21:06] some surface mount component in the
[21:08] ELSFP uh, discrete laser,
[21:11] the improvement uh
[21:13] the reliability or the MTBF raises
[21:15] dramatically.
[21:17] Okay, so so you would say fundamentally
[21:19] that that issue with the ELSFP module
[21:22] has nothing to do with CPU reliability
[21:24] and if you fix that, you should be able
[21:25] to get something more like a 10x
[21:27] improvement than a 2x improvement, which
[21:29] seems a little bit more than a smaller
[21:30] bag of parts, but but this is where we
[21:32] are at the moment.
[21:33] If we talk about non-serviceable errors,
[21:36] this goes back to the the the issue of
[21:38] reliability, availability, and
[21:39] serviceability, right? These are These
[21:41] are non-serviceable issues, which means
[21:43] that the entire box has to be pulled.
[21:45] The entire CPO switch box has to be
[21:48] pulled if one of these things happened.
[21:50] It's over 20 million, so it's two and a
[21:53] half times better than the serviceable
[21:55] error.
[21:56] Uh again, this is an important
[21:57] consideration for us because it goes
[21:59] back to how much excess capacity you
[22:00] have to put into the fleet to account
[22:02] for the fact that these things are down.
[22:05] Maybe I won't because Jose's waving at
[22:07] me. I won't say too much about why we
[22:08] think that CPO is more reliable. Smaller
[22:10] bag of parts sort of matters.
[22:12] Um deeper levels of integration and
[22:14] design in the manufacturing process
[22:17] matters
[22:18] I think an awful lot. I think there's
[22:19] some optimization which is going on
[22:21] here. Reduced levels of intervention,
[22:23] all that sort of matters.
[22:25] I mean, you know, we think the first
[22:27] field serviceable ELSFP module helps a
[22:29] lot driving this of sort of availability
[22:32] and serviceability part of the RAS
[22:34] question, which is a significant part
[22:35] for that.
[22:37] But this is also an experiment which is
[22:40] maybe halfway through, maybe less,
[22:42] right? We have 50 million device hours
[22:45] on the Bailey
[22:47] um phase two parts and we can't report
[22:49] an MTBF with any level of confidence
[22:51] simply because we're not getting errors.
[22:54] They're not sort of creating for us the
[22:56] sorts of signals that we're creating and
[22:57] calling errors at this point. So we've
[23:00] got some more data to get there.
[23:01] Hopefully we'll get to the point where
[23:02] we have enough failures that we can
[23:03] actually give you a confident MTBF and
[23:05] sort of say this is where we think the
[23:07] gold standard or at least a baseline
[23:08] should be that all future systems should
[23:10] exceed.
[23:11] But this is uh in the end a good result
[23:14] even if it's a null result.
[23:16] Others have talked about OCI MSA. I
[23:20] I'd love to talk to you about this in
[23:21] detail, but but because I'm going to
[23:23] break Jose's heart if I do that, I
[23:24] won't. Please go uh download the
[23:26] standard. Uh we'd love to have you adopt
[23:29] this and start
[23:31] fabricating parts and and compliant
[23:33] parts here so we can take vendor A's
[23:35] parts and connect them to vendor B's
[23:36] parts. I really think that this is going
[23:37] to be a significant part of what enables
[23:40] optical uh scale-up domains moving
[23:42] forward. We can sort of break from the
[23:44] tyranny maybe not of copper in this
[23:45] case, but at least serdes and the power
[23:47] associated with that.
[23:49] And I guess I will say lastly uh scale
[23:52] larger scale-up domains are coming. It's
[23:54] going to be harder and harder to do this
[23:55] in the electrical realm with the sort of
[23:58] racks that we're talking about and the
[23:59] scale and the implications of that. CPO
[24:01] and OCI really are going to help with
[24:03] sort of um scale-up domains optical
[24:06] scale-up domains I hope in the future.
[24:08] Uh the results that we're getting on
[24:10] reliability are encouraging. Experiments
[24:13] now finished. I promise we'll come back
[24:14] 6 months or so and sort of give you a
[24:17] sort of result from there, but we're
[24:18] hoping that the sort of this data and
[24:21] and sort of coalescing around the OCI
[24:23] MSA will help us get there as an
[24:24] industry where we can, you know,
[24:27] make the world a better place with
[24:28] optical scale-up. Thank you.
[24:33] All right. Thanks a lot, Drew.
[24:36] I am uh I'm Dave Lazovsky. I'm uh the uh
[24:40] EVP and general manager of the data
[24:43] center networking group at Marvell. And
[24:45] I'll start by saying that I'm I'm
[24:47] thrilled to be at Marvell. It's nice uh
[24:49] not just to have uh the reinforcements
[24:52] uh the you know, 5 6 7x the number of
[24:54] resources available within our group, uh
[24:57] but uh to have the breadth uh and depth
[25:00] of technical acumen across the team that
[25:01] we've got, the depth uh of partnerships
[25:04] that we've got across the industry with
[25:06] uh our our hyperscale customers, the GPU
[25:08] manufacturers, and with the overall
[25:10] ecosystem with all of you. Um, but also
[25:12] uh to have uh access to building system
[25:14] level solutions, uh which is something
[25:17] that I'm going to share with you a bit
[25:19] today.
[25:20] Um, here's uh
[25:21] our forward-looking statement, so so go
[25:23] ahead and grab a quick picture.
[25:25] Um, so building on what uh
[25:28] was I think
[25:29] uh included in a large part of what you
[25:31] heard this morning if you were uh
[25:33] joining uh the uh the hyperscaler talk
[25:36] at the beginning of the day and and uh
[25:37] Richard Ho's um uh keynote this morning.
[25:41] Um, there's a tremendous amount of
[25:42] innovation taking place at a cadence
[25:44] that's just unprecedented. Um, and uh
[25:47] all of these innovations across the uh
[25:50] artificial intelligence workloads are
[25:52] having a direct impact as we know on the
[25:54] data center infrastructure that is
[25:56] supporting them. Um, uh each uh from the
[26:00] from the development of uh large
[26:01] foundation models to just 12 months ago
[26:05] the concept of uh reasoning and
[26:07] first-time compute was a new thing. Uh
[26:09] now, if you don't have uh uh reasoning
[26:11] models in your uh bailiwick, um and you
[26:14] don't have the infrastructure including
[26:16] the ability to deal with very large KB
[26:18] caches that are that are increasing by
[26:20] an order of magnitude um literally about
[26:22] every 6 months, um it becomes really
[26:25] challenging, right, to be and maintain
[26:27] yourself at the leading edge of of
[26:28] capability infrastructure. Um, agentic
[26:31] AI workloads, mixtures of experts,
[26:34] putting more and more and more strain on
[26:37] developing system level solutions,
[26:39] right? Um, compute uh is and remains
[26:41] important as a prerequisite for
[26:43] capability, uh but network connectivity
[26:46] um uh across
[26:48] uh the uh the entire network scale up,
[26:51] scale out, scale across is is we know
[26:54] and we've heard repeated times today
[26:55] critical. All of that leading then to
[26:57] the to the point where ultimately
[26:59] um these workloads need to not just be
[27:01] performant, uh but they need to be uh
[27:03] profitable, right? So, it's uh getting
[27:05] to the point where we've got economies
[27:06] of scale and we're able to tap into the
[27:09] the uh return element of the ROI
[27:11] equation in artificial intelligence,
[27:13] which is inference, right? So, training
[27:15] in large part uh is the investment that
[27:17] takes place. So, a lot a lot of the
[27:18] infrastructure that will be deployed
[27:20] over the course of the balance of this
[27:21] decade will be inference-specific
[27:23] infrastructure.
[27:25] Um we also heard from the panelists this
[27:28] morning about uh the implications on
[27:31] energy and the importance of energy. Um
[27:33] and it's um
[27:35] it's it's tricky, right, to uh to
[27:37] decouple the implications of uh how much
[27:40] power is being consumed by computation
[27:43] versus communication. And it honestly is
[27:45] in large part a function of as Hong
[27:47] pointed out earlier today, um it's your
[27:49] architecture, which can play a large
[27:51] role in this. Um
[27:54] uh it is also uh the the workloads that
[27:57] you're running, right? So, many of our
[27:58] customers have artificial intelligence
[28:00] workloads running in extremely
[28:02] high-volume manufacturing that are
[28:03] seeing about basically 25 to 30% model
[28:08] flop utilization. Right? So, the other
[28:11] roughly 70% of the time uh these systems
[28:14] are waiting for
[28:16] um
[28:17] uh memory transactions to take place
[28:18] over the scale-up network. Right? So,
[28:20] it's not terribly efficient. The other
[28:22] thing that hides a bit um of the the
[28:25] total power consumption in these systems
[28:27] uh is the fact that the SerDes is built
[28:30] into the XPU, right? And so, almost all
[28:34] of the scale-up network uh data
[28:36] transmission that's taking place is
[28:38] taking place over a SerDes on both the
[28:40] XPU and the switch side. So, it's
[28:41] difficult to deconvolve the
[28:43] contributions to total system power deal
[28:45] total data transmission. So, these are
[28:47] areas of focus as we move increasingly
[28:50] toward optical interconnectivity
[28:52] to
[28:53] to significantly reduce the total amount
[28:55] of power consumption. So,
[28:57] depending on the workload, depending on
[28:58] the network configuration, north of 50%
[29:01] of power consumption for certain
[29:02] workloads
[29:03] can be can be moving information around,
[29:06] not just the compute.
[29:08] So,
[29:09] again, one of the great things about
[29:10] being at Marvell is that
[29:13] we have a broad suite of optical
[29:15] interconnectivity capabilities, and we
[29:17] also
[29:18] are
[29:19] leaders in 224 gig and really advanced
[29:22] serdes technology for copper.
[29:24] So, it gives us the ability to develop a
[29:27] variety of interconnectivity solutions
[29:29] for our customers to meet precisely
[29:30] their needs. The other thing that we're
[29:32] doing is partnering very closely with
[29:34] them
[29:35] with across the the ecosystem with the
[29:38] hyperscalers working backward, which
[29:40] I'll share with you,
[29:41] to develop solutions that meet not just
[29:44] the XP requirements for custom custom
[29:46] silicon,
[29:47] but we're developing custom networking
[29:50] solutions to make sure that their entire
[29:52] needs are being met. And we've got a
[29:54] portfolio of not just copper
[29:56] interconnectivity capabilities, but
[29:58] optical interconnectivity capabilities
[30:01] going from taking the light engines that
[30:03] we're using for pluggable data center
[30:05] transceivers and moving them to near
[30:06] package optics implementations. We are
[30:10] working
[30:11] in the qualification
[30:14] deeper levels of conventional MSA
[30:17] compatible
[30:19] co-packaged optics and near package
[30:21] optics.
[30:22] And then we have the deep the deepest
[30:24] level of optical interconnectivity
[30:25] that's that's capable, which is the
[30:27] photonic fabric, which
[30:29] has a series of incremental advantages.
[30:32] We heard over and over this morning, so
[30:35] we're not going to beat this horse on
[30:38] scale-up network connectivity. That's
[30:39] what we're focused on. But I think as we
[30:42] we did her also, the the requirements
[30:44] here are entirely different, right? So,
[30:45] it's not just 10x the bandwidth, and the
[30:47] advantage being the reach, right?
[30:50] Because there is one company who's done
[30:52] a really amazing job of commercializing
[30:55] with extremely high volume scale-up
[30:58] network solutions today, not just the
[30:59] link, but the switch. And that company
[31:02] is led by a guy named Jensen Huang,
[31:04] right? So, NVLink and NVLink Switch is
[31:06] dominant clearly in scale-up network
[31:08] connectivity. As all of the hyperscalers
[31:10] have, um, uh,
[31:12] in in in, you know, in particular, uh,
[31:15] Google, um, but, uh, Annapurna, those
[31:18] that are um, working on, uh, Meta,
[31:21] Microsoft, that are working on more
[31:22] comprehensive custom silicon solutions
[31:25] that are not again just the XPU
[31:26] development, uh, for, um,
[31:29] training and inference accelerators, but
[31:30] custom network connectivity, they're all
[31:32] working on developing their own versions
[31:34] of, uh, scale-up connectivity that will
[31:36] evolve from copper, uh, to, uh, optical
[31:40] inner connectivity over the course next
[31:42] few years. Some of them, um, are moving
[31:44] a lot faster in this transition, uh,
[31:47] than, uh,
[31:48] you know, clearly a couple of the of the
[31:49] folks, uh, that, uh, that we heard from
[31:51] today. Um, and it's that initial
[31:54] transition is going to be extremely
[31:56] important, right? To get to the point
[31:58] where we can build on the kind of data
[32:00] that Drew just pointed out, uh, earlier
[32:02] today that that Meta has done, looking
[32:04] at the inherent advantages of, um,
[32:07] simpler,
[32:09] more integrated optical inner
[32:11] connectivity solutions. All right? I
[32:13] think we learned, so, for those of us
[32:14] that, uh, studied engineering, which is
[32:16] probably 95% of this room,
[32:19] uh, that, uh, keeping things simple, uh,
[32:21] goes a long way toward reliability, um,
[32:24] and, uh, and serviceability and
[32:26] availability.
[32:27] So, um, in addition to these metrics
[32:30] that are shown here with reach, power,
[32:31] latency being less than on the order of
[32:33] 200 nanoseconds, um, yeah, power's
[32:36] really really critical
[32:38] um and larger and larger uh scale-up
[32:42] domains which require more and more
[32:44] reach and um
[32:46] uh and and also the radix uh for these
[32:48] uh systems will continue to uh to
[32:51] increase. The last metric here is cost,
[32:54] right? We need to have solutions that
[32:56] are compatible or in the same general
[32:57] ballpark as copper uh copper
[32:59] interconnectivity uh for uh higher
[33:02] performance scale-up network solutions
[33:03] which uh depending on the scale-up
[33:06] network they're using is on the order of
[33:07] 5 cents per gigabit per second.
[33:09] So, um from the standpoint of the
[33:11] photonic fabric, which is what we
[33:12] developed at Celestial AI that was
[33:14] acquired by Marvell, we've got a series
[33:16] of innovations that took place um that
[33:18] we focused on to drive differentiation
[33:20] for really what should be an ultimate
[33:22] solution. Um again, some customers are
[33:24] moving right to it. Many others will go
[33:27] through multiple iterations of uh
[33:29] uh copper to NPO to CPU to finally get
[33:32] there. All right. So, um uh it starts
[33:35] here with uh full optimization of the
[33:37] metrics that we talked about before on
[33:39] the previous slide, which is uh getting
[33:41] to power optimized systems, which is
[33:43] eliminating DSPs and moving directly to
[33:45] analog serdes um and providing an an
[33:47] ecosystem that is that is compatible
[33:50] with things like um
[33:52] uh
[33:52] uh advanced packaging uh TSMC CoWoS uh
[33:56] SRNL and then ensuring that there's an a
[33:59] high-volume manufacturing supply chain
[34:01] to support needs.
[34:02] What we've done is we've established a
[34:04] playbook um with our leading hyperscale
[34:07] partners to make this transition easy,
[34:09] um
[34:10] lower risk than if they're just to go
[34:12] jump into a a CPO solution.
[34:15] How we're doing this is providing
[34:17] protocol compatibility and physical
[34:20] interface physical layer compatibility
[34:21] leveraging um some established standards
[34:25] that are taking place with UAL, uh even
[34:27] EASON uh for a scale-up network
[34:30] connectivity protocols. And then um the
[34:32] emergence of standards uh uh for
[34:35] physical interfaces like UCIE, which
[34:37] provides us the ability, which is shown
[34:38] here, to go from without changing at all
[34:41] the XPU core chip nor the switch core
[34:45] chip, we can move from a co-packaged
[34:47] copper interface to a co-packaged
[34:50] optics, a photonic fabric-based
[34:52] interface um that is completely software
[34:54] compatible, and it allows our customers
[34:56] to get a cycle of learning in very
[34:58] high-volume manufacturing before they
[34:59] really start on the gas and scale to to
[35:01] uh to CPO for the dominant scale of
[35:03] network solution. Um this uh is an
[35:06] approach that we think can significantly
[35:07] reduce risk for the implementation of
[35:10] systems. So,
[35:12] um again, one of the benefits of uh of
[35:15] being inside of Marvell is having a full
[35:17] toolkit of not just um the optical inner
[35:20] connectivity, but the switch uh working
[35:22] on scale-out scale-up switches, which
[35:24] now falls in into my group, so we're
[35:25] working really closely connecting the
[35:28] the dots um and ensuring that we have
[35:30] the full link capability um particularly
[35:32] for scale-up network connectivity from
[35:34] XPU working with our custom silicon
[35:36] partners to um the scale-up switch, um
[35:41] and then making sure also that we have a
[35:43] roadmap that facilitates the ability
[35:45] just to say yes to customers, especially
[35:47] the largest of customers, to uh tailor
[35:50] solutions to meet their precisely their
[35:52] their requirements. And then as we move
[35:54] toward larger scale-up domains, again,
[35:57] um optics facilitates the ability to
[36:00] move outside the limitations of a rack.
[36:02] Um this has layers of benefit that we
[36:03] didn't fully appreciate that are hard to
[36:05] go into when I'm out of time, uh but
[36:07] it's something uh something worth
[36:08] talking about. Also, the ability to tap
[36:11] into memory disaggregation, like you
[36:13] heard from uh both Microsoft and from uh
[36:16] Richard Hong this morning, the ability
[36:18] to scale memory capacity and bandwidth
[36:21] independent from compute will become
[36:23] increasingly
[36:24] needed um as we progress toward more
[36:27] advanced workloads. The objective here
[36:29] is future proofing systems and data
[36:31] center infrastructure because you can't
[36:33] possibly innovate at the rate of
[36:36] innovation that's taking place with
[36:38] today's advanced AI workloads. And then
[36:40] taking these scale-up domains and
[36:44] interconnecting multiple scale-up
[36:46] domains with scale out. All right, so
[36:48] again, we have the
[36:50] just the true benefit at Marvell of
[36:52] having all the tools in the toolbox to
[36:54] work with customers to deliver system
[36:57] level solutions.
[36:58] And increasingly we'll be delivering
[37:00] system level solutions meeting customer
[37:01] needs across all of the aspects of
[37:04] interconnectivity.
[37:06] So,
[37:08] it is a it's a it's an approach that we
[37:10] take
[37:12] in deep collaborations with customers
[37:14] working backward from the algorithms and
[37:16] the workloads to determine what
[37:18] artificial intelligence infrastructure
[37:19] was required. And because there's such
[37:21] concentration in this market, we're able
[37:23] to just tailor solutions to meet meet
[37:25] our customers' needs.
[37:27] So, it is beneficial to have not just
[37:31] the link or the switch or custom silicon
[37:34] operating independently.
[37:37] The
[37:38] objective for us is is to come to our
[37:43] customer base with really comprehensive
[37:45] solutions and deep, broad, and deep and
[37:47] broad partnerships
[37:49] that can deliver a level of integration
[37:51] for their near-term requirements and
[37:53] their long-term road maps across all of
[37:56] their system needs. Deeper partnerships
[37:59] than what I think this industry has
[38:01] experienced
[38:03] today.
[38:05] So, with that,
[38:06] again, thrilled to thrilled to be a part
[38:08] of this team leveraging the
[38:11] the extensive history, the tens of
[38:13] millions of
[38:15] optical product units that Marvell has
[38:17] delivered. We got we have not just one
[38:19] of the industry's largest silicon
[38:21] photonics and optical teams, but the
[38:24] team is just unbelievable. Just the the
[38:26] the depth and breadth of technical
[38:27] acumen is
[38:28] is is available to our our leading
[38:30] partners and the this the strength of
[38:31] the supply chain with many of you.
[38:34] We're we're we're thrilled to have you
[38:36] as partners and the depth again of our
[38:38] customer relationships also with with
[38:40] many of you.
[38:42] We couldn't ask for better partners to
[38:43] have. So with that
[38:45] I'll I'll hand it over.
[38:48] All right. Good afternoon.
[38:50] I'm going to talk about open CPX
[38:54] co-packaged and near-packaged mixed
[38:56] media interface.
[38:58] So first of all, we really need to need
[39:00] to acknowledge that high-speed 30s have
[39:02] been the currency of the data center
[39:04] ever since there have been data centers
[39:06] really and that goes back quarter of a
[39:08] century as you see in the chart
[39:11] on the bottom of the slide. So switch
[39:13] capacity has been scaling very nicely
[39:16] exponentially at 40% per year.
[39:19] Per lane speed of 30s has been very
[39:21] consistently scaling at 20% per year. We
[39:24] are at 200 gig today. We will be going
[39:27] to 400 gig. There's very good progress
[39:29] in the industry. It won't happen next
[39:32] year, but it doesn't have to happen next
[39:33] year. If you look at the exponential
[39:34] scaling, if it happens by 2030, that's
[39:37] plenty. That's where we need to be.
[39:39] And why are high-speed 30s so important?
[39:42] That's because that's the the only way
[39:45] to escape packages if you believe that
[39:48] copper is a life.
[39:50] That's the only way is to go high-speed
[39:52] because otherwise you need too many
[39:53] bumps. And then once you are outside the
[39:56] package, you can go over PCB traces. You
[39:59] can go over flyover cables. You can even
[40:02] use what Broadcom introduced at OCP last
[40:05] year over the integrated copper attach
[40:08] or ICA
[40:09] through a second level connector through
[40:13] to to either near-package or co-package,
[40:16] however you want to call it, doesn't
[40:17] matter,
[40:18] um copper or optics.
[40:21] Now, people are confused. I heard that
[40:23] in many uh many questions today. What's
[40:26] There are these three MSAs out there
[40:28] now, XPO, Open CPX, OCI. What's the
[40:31] difference? And it's really very very
[40:33] very simple. It's You only have to ask
[40:35] yourself, is copper alive, yes or no?
[40:38] That's the only question you need to ask
[40:40] yourself.
[40:41] If you believe that copper will be used,
[40:44] then you go XPO or Open CPX, because
[40:47] then you need high-speed serdes, that's
[40:49] the only way to escape the package, and
[40:51] then you can use either copper or
[40:53] optics. If you believe that copper is
[40:55] dead, meaning really really dead, and
[40:58] you don't have an alternative anymore,
[41:00] then you use OCI. You need to integrate
[41:03] the chiplet onto uh your package, but
[41:07] that's it. Then you can go slow and
[41:09] wide, that's fine, but don't even think
[41:11] of using a PCB trace anymore. That's
[41:13] dead, okay?
[41:15] So, um let's look a little bit about uh
[41:18] on AI uh clusters, and especially scale
[41:21] up, so since that's a scale up panel. Um
[41:25] the limit, the size, the number of GPUs
[41:28] that you can connect in your scale up uh
[41:31] cluster, is limited by the radix of your
[41:34] switch.
[41:35] So, that's the fundamental limit. That's
[41:37] what your scale up cluster is limited
[41:39] by. And um
[41:42] what the the the name of the game now is
[41:44] to get as many GPUs into the copper
[41:47] radius, as Drew called it. That's a very
[41:50] nice word. I I'm going to use it from
[41:52] now on. The copper radius of your
[41:54] switch.
[41:55] So, obviously, at 200G with passive
[41:57] copper, that's not a very large radius.
[41:59] That's a meter,
[42:01] give or take. But, with active copper,
[42:04] for just a picojoule and a half,
[42:07] you can get to 4 and 1/2 m.
[42:10] And that's a pretty large radius. At in
[42:12] four, four and a half meters, you can
[42:14] get to a thousand GPUs or scale up
[42:16] cluster. That's possible.
[42:18] And now you're there at one scale up
[42:21] domain of a thousand GPUs. How do you
[42:23] get to more GPUs? Scale out? Well, scale
[42:27] out has an order of magnitude less
[42:28] bandwidth. It's not ideal. You would
[42:31] want ideally to go to a much larger
[42:34] scale up clusters if you can. The only
[42:36] way to do that because of the radix
[42:38] limitation is to go to a multi-tier
[42:41] scale up network.
[42:43] If you do that, then in each
[42:46] scale up tier one cluster, you can now
[42:49] only accommodate half the number of GPUs
[42:50] because the switches point to the second
[42:53] tier for the for the upstream. And what
[42:55] you see here is really that you can make
[42:58] this work. The the pink
[43:00] rack cluster, if you want, could be like
[43:03] three or four racks. You can make this
[43:05] work with active copper, no problem. And
[43:07] then of course, to go to the tier two
[43:08] switches, you need optics because they
[43:10] are going to be far away. But what you
[43:12] see is that those red switches in the
[43:14] tier one, they are mixed media. They
[43:16] speak copper upwards and optics
[43:18] downwards.
[43:20] You can also slice the cat in different
[43:21] way and
[43:24] have all the GPUs with optics, fine. You
[43:26] can have that and then go to a mega
[43:29] radix switch that will then use copper
[43:32] and optics inside just to save power cuz
[43:35] you use copper when you can cuz it's
[43:37] most reliable, lowest power, and
[43:39] cheapest. So, but even here, you're
[43:42] talking mixed media. You're now talking
[43:43] mixed media inside the giant two-stage
[43:46] switch. So, mixed media is just the name
[43:48] of the game if you want to build low
[43:51] power solutions. And just to visualize
[43:53] that a little bit more,
[43:55] what we used to have five years ago was
[43:58] a choice between passive copper and
[44:00] re-timed optics. One goes hundreds of
[44:03] meters at 25 watts per 1.6T today. The
[44:06] other costs you 0 W
[44:08] once you have the 30s and you can only
[44:10] go 1 m for 200 G.
[44:13] But now there are new kids on the block.
[44:15] Now we have more tools in our toolbox to
[44:17] really power optimize our systems.
[44:20] Coming from the bottom, we use active
[44:22] copper cable. Again, with a picojoule
[44:24] and a half, you now go not 1 m but 4 m
[44:28] or 4 and 1/2 m. Or coming from the top,
[44:31] instead of using re-timed optics, you
[44:33] use half re-timed or linear optics and
[44:35] you go
[44:37] you go the same distance hundreds of
[44:38] meters with 10 W.
[44:40] So then it's really about avoiding
[44:42] re-timers but not avoiding electronics.
[44:44] Analog ACCs are totally fine and they
[44:47] really increase your reach, your copper
[44:50] reach, your copper radius to where it
[44:52] needs to be. It doesn't need to have to
[44:54] be larger than the the radius of your
[44:56] switch. That's it doesn't need to be
[44:58] larger than that. That's all you need to
[44:59] do.
[45:01] Um now, why then co-packaging and not
[45:04] front panel pluggables?
[45:06] There is one reason for that and that's
[45:08] density saves power. Why why is that? So
[45:12] if you compare front panel pluggable
[45:13] architecture with a co-packaged
[45:16] architecture, you see that the front
[45:18] panel architecture has longer copper
[45:20] traces or could be flyover cables but
[45:22] it's still longer electrical traces to
[45:26] the the optics. And that's more loss.
[45:28] You might need a re-timer or a half
[45:30] re-time solution and that costs you
[45:33] power. But that's not that's the obvious
[45:35] density argument but not the only one.
[45:38] There is another one and that's if you
[45:40] look at the front panel of a front panel
[45:43] pluggable box and you build it with
[45:45] OSFPs, you can fit 32 OSFPs on that
[45:49] front panel in a 1 RU box. And that's
[45:52] just 50 terabit worth of escape.
[45:55] But with the and that was totally fine
[45:57] as long as thermal
[46:00] things were air cooled. But now things
[46:03] are liquid cooled, so you can cram many
[46:05] more ASICs into this one RU tray than
[46:07] you could previously.
[46:09] But what does that mean? You now need
[46:10] much more escape bandwidth on your front
[46:13] panel on your one RU.
[46:15] And in order to achieve that, you need
[46:17] to densify your one RU face plate. One
[46:22] way to do that is XPO, a much denser
[46:25] form factor that India I'm sure will
[46:26] talk about in much more detail. Another
[46:28] one another way to do that is CPO.
[46:31] Because now your front panel consists of
[46:34] these very nice MMC or SNMT, these very
[46:37] nice miniature fiber array connectors.
[46:40] And now you have an order of magnitude
[46:41] more escape bandwidth.
[46:43] 10 times more than with with pluggables.
[46:47] So this is these dense 12.8T pluggables,
[46:51] XPO, the new pluggable form factor, and
[46:54] that's a version that we're building for
[46:56] LPO and RTLR.
[46:59] They have they're very very dense. They
[47:01] allow you to do a thousand lanes in a
[47:04] one OU unit.
[47:06] And for
[47:08] co-packaging, this is CPX. This is a
[47:11] mixed media
[47:12] co-packaged socket
[47:15] that allows you to put both
[47:17] co-packaged copper and co-packaged
[47:20] optics into the same socket. It's a 6.4T
[47:23] connector. You can use it co-packaged,
[47:25] you can use it near packaged. It doesn't
[47:27] require any compression hardware. No
[47:29] nuts and bolts. Just plug it in and
[47:31] done. And it scales to 1024 lanes. It
[47:34] scales to 400G per lane.
[47:37] And last week we announced the formation
[47:40] of the Open CPX MSA with the founding
[47:43] members Ciena, Coherent, Marvell, Molex,
[47:45] Samtec, and Terahop. You can go to the
[47:48] website to learn more.
[47:50] And you can now build also the ICA with
[47:54] that. You can I to this two-level
[47:57] interface from the package, which will
[47:59] be nice to standardize a package
[48:00] interface as well. So, you go out from
[48:03] the package onto the ICA and then out
[48:06] through the Open CPX module.
[48:09] And that's what the Open CPX module
[48:11] looks. That's actually very large. It
[48:13] always looks so large on the slide. So,
[48:15] this is what it really looks like. So,
[48:17] this is the Open CPX module.
[48:20] And 6.4 terabit fully pluggable, low
[48:24] power, no compression hardware. What
[48:26] else would you want? Super high density.
[48:29] Thank you.
[48:33] Okay, so it's admittedly a little
[48:35] difficult to tell you anything new after
[48:38] all the things we already heard today.
[48:40] So, I will try my best to focus on
[48:41] things that that may not have been said
[48:44] earlier, but you know. So, let me start
[48:47] with
[48:49] a thesis that we're actually
[48:51] under forecasting the need for future
[48:54] AI connectivity, meaning optics.
[48:58] So, the there's sort of three columns
[49:00] here sorted by scale of bandwidth per
[49:03] per XPU or GPU and then scale out
[49:05] bandwidth.
[49:06] And the left column is kind of where we
[49:08] were yesterday.
[49:10] The middle column is kind of where we're
[49:12] this year. And the right column is where
[49:14] we probably are in 28.
[49:16] So, if you multiply it and then there's
[49:18] a number of GPUs or XPUs per data
[49:20] center.
[49:21] So, if you multiply this all out and
[49:24] look for what for two years,
[49:26] the calculation is that a million GPU
[49:29] data center would need 8 million scale
[49:32] out optics and 128 million scale up
[49:36] optics.
[49:37] And that's maybe 10% of the industry,
[49:40] right? So, if there's 10 million GPUs
[49:41] per year, just assume it would be over 1
[49:45] billion optics, 600 gig equivalents,
[49:48] right? So, you heard it here first.
[49:50] Now, obviously, we
[49:53] it's whatever we have to do isn't good
[49:54] enough. We need much lower failure
[49:56] rates. Every fit matters. We need much
[49:59] lower power. Every picojoule matters.
[50:02] And we need much higher density. Every
[50:04] act unit matters. And each of these
[50:06] metrics is of course multiplied by
[50:09] millions of equivalent units.
[50:11] But there's one thing we need more than
[50:12] anything, which is
[50:14] the ability to
[50:16] ship 1 billion units per year.
[50:20] These are again 600 gig equivalent
[50:21] units, which is primarily driven by the
[50:24] scale up market, which is 10 times the
[50:25] volume of the scale out market, right?
[50:29] So, this is a list of, you know,
[50:31] well-known technologies people have been
[50:32] working on and are working on, in
[50:35] particular for, you know, for all of the
[50:37] above, scale out, scale up, scale out,
[50:39] and scale out.
[50:40] In scale up today, it's mostly copper
[50:43] cables, right? And you will use copper
[50:45] as long as possible.
[50:46] In you have active copper, you can go 4
[50:49] 5 m. It's still very low picojoule per
[50:51] bit.
[50:52] Very low failure rates. There's a lot of
[50:54] excitement around RF microwave, which is
[50:57] a, you know, we heard earlier, it's a
[50:58] non-photonics technology that can go
[51:01] through a waveguide, you know, 10 or 20
[51:03] m.
[51:04] Again, at very low picojoule per bit.
[51:06] And then we're getting to the DR8 and
[51:08] the single mode optics. So, there's
[51:10] actually two types of
[51:12] single mode optics, the call it the low
[51:14] power variants, which are not silicon
[51:17] photonics, like indium phosphate, indium
[51:19] light bit. And the higher power
[51:21] variants, which I highlighted in the
[51:23] yellow because that's kind of the
[51:24] dominant focus of all the
[51:26] uh CPO uh and focus today, where this is
[51:29] yellow line.
[51:30] And then beyond that, you know, we have
[51:32] FR4, LR4, coherent light, and uh CR uh
[51:37] full long long reach.
[51:39] Now,
[51:41] the point of this slide is there's what,
[51:43] 10 or nine technologies on this list
[51:45] that are all relevant to
[51:48] AI data centers. You know, that's it's
[51:50] not the case that one technology solves
[51:53] all problems, right? And obviously, as
[51:55] you go down this list, the failure
[51:57] rates, unfortunately, will increase and
[51:59] the cost also goes up a bit.
[52:02] So,
[52:03] to scale from, call it a hundred million
[52:06] this year to a billion,
[52:08] you know, is a
[52:10] is a challenge. We have 2 years to get
[52:12] ready for this.
[52:13] Um it probably can be done, you know,
[52:15] the people are struggling with making
[52:16] more lasers, more optical isolators,
[52:18] more indium phosphide supply, you know,
[52:21] you name it.
[52:22] Uh but again, with time and investment,
[52:24] this can be solved. Um in fact, uh
[52:27] mentioned earlier today, scale-up is
[52:28] mostly copper cables. People had trouble
[52:31] building billions of copper cables,
[52:32] right? And they're still working on
[52:33] that. But transitioning scale-up to
[52:35] optics is in fact the biggest challenge.
[52:38] And my main point here is that the
[52:40] easiest technology and the lowest power
[52:42] and the most reliable for scale-up is
[52:45] not optics, right? It's copper cables
[52:47] and short-range microwave. So, um
[52:50] so, I think that's the grand challenge
[52:52] here. And let me just
[52:54] attempt to have one slide about CPU
[52:57] here.
[52:58] HVM means high volume manufacturing and
[53:01] clearly, you know, it takes time,
[53:03] effort, and lots of resources to make
[53:05] this happen. Um a very large company is
[53:07] spending a fortune to make this happen
[53:10] from the foundry to the packaging to the
[53:12] OSAT, the FAUs, the ELSs, the system to
[53:15] all of this has to scale by tremendous
[53:17] amounts.
[53:18] And the key issue here is that you
[53:20] cannot ship the GPU or the switch chip
[53:23] until the associated CPU is ready for
[53:26] high volume. It's kind of a chicken and
[53:28] egg problem. Like, ideally, we'd decide
[53:30] to put the CPU on the switch when you
[53:32] know it's really working because if it's
[53:34] not working, you can't ship and that
[53:35] would be really bad. So, generally
[53:37] speaking, and I'm really generalizing
[53:39] here,
[53:40] uh from a sort of MSA spec to high
[53:43] volume, call it roughly 2 years.
[53:46] So, and this is true by the way for
[53:47] conventional optics as well. So, I'm not
[53:49] This is not a special commodity issue.
[53:51] Um, but
[53:53] if you started today, we need a really
[53:55] spec to intercept the 400 gig per lane
[53:57] generation.
[53:58] And that means you need the spec today
[53:59] to be ready in high volume in '28 for
[54:02] that intercept. And, you know, perhaps
[54:04] that can be done. In fact, my my
[54:06] personal opinion is that the volume
[54:08] intercept point for CPUs actually 400
[54:09] gig per lane with 400 gig per lane sort
[54:12] of and and whatever optics that match
[54:14] that.
[54:15] Um, but going back to what we do in the
[54:17] meanwhile, you know, um,
[54:20] we launched OCP actually 10 years ago in
[54:22] 2016 uh, together with Google we formed
[54:25] the OCP MSA and it it has been a
[54:28] stunning success. More than 100 million
[54:30] units will ship this year. It supports
[54:32] any kind of optics, copper cables, DR4,
[54:34] LR4, SR4s,
[54:36] current light and whatever comes next.
[54:38] And one reason it was so useful, it
[54:40] could transition from 400 gig to 800 gig
[54:42] to 600 gig with a robust thermal
[54:44] envelope between 30 40 watts air cooling
[54:47] and a nice 32 ports per 1U front panel
[54:49] density. So, that was great, but it's
[54:52] not good enough because we actually hit
[54:54] the what I call the thermal envelope
[54:56] density limit. We cannot get more optics
[54:59] in this box because the thermal stones
[55:01] support it.
[55:02] What we need is something that's liquid
[55:04] cooled and it's much denser and as you
[55:05] know, we we call this the XPO. It's the
[55:07] equivalent of
[55:09] eight 600 gig OCPs wrapped into a much
[55:12] denser form factor. And, uh, again, this
[55:15] allows us to build a future 240 switch
[55:18] in 1 OU. And this contrasts with the
[55:21] today's, you know, 600 gig OCP solution
[55:24] that would require 4 U.
[55:27] So, I just want to show you the dramatic
[55:29] improvement here. So, assuming for a
[55:31] second 512 GPUs, each one of those 25.6
[55:34] T scale up and, you know, optics
[55:36] connection to OCP switch racks, you
[55:39] would need eight OCP switch racks to
[55:42] connect four GPU racks, each one having
[55:45] 128 GPUs. It doesn't make sense, right?
[55:48] With XPU, it looks like this. It's half
[55:51] the footprint in, you know, for the same
[55:53] thing.
[55:54] Now, at the data center level, this this
[55:56] the saving is so much larger, of course.
[55:58] This is an actual design a customer
[56:00] design point of a 400 MW data center
[56:02] with 128,000
[56:04] XPUs. So, or GPUs, this 128 per rack is
[56:08] 1024 XPU racks shown in blue. Today, and
[56:12] this is for both scale up and scale out,
[56:14] and this is assuming only 12.8T scale
[56:17] up, so it's a less less than the
[56:18] previous slide.
[56:20] You would need 1400
[56:22] switch racks to connect a thousand GPU
[56:25] racks. Again, it doesn't make a lot of
[56:28] sense. With XPU, it looks like this.
[56:32] You're saving over 1,000 switch racks.
[56:35] This is not quite half the total,
[56:36] because the GPU racks are still there,
[56:38] but you're saving 44% of the floor
[56:40] space. You know, all these buildings
[56:42] that are being built for billions and
[56:44] millions of dollars, they could have
[56:45] built them half the size.
[56:47] So, having a
[56:49] 75% reduction in switch racks is
[56:52] actually highly desirable, because
[56:55] basically the the sheet metal, the
[56:58] the racks, you know, the steel, the bus
[56:59] bars, the manifolds is a tax on the
[57:02] payload and ultimately on the revenue
[57:04] generation of the data center, right?
[57:06] You want fewer racks, you don't want
[57:07] more racks.
[57:08] Also, saving almost half the footprint
[57:10] is highly desirable. It will save you
[57:12] billions of dollars.
[57:14] There's another benefit. The wires are
[57:16] getting shorter. So, some of these
[57:18] lowest power technologies work much
[57:20] better with shorter connections than the
[57:22] longer ones. And again, with dense
[57:25] connectivity, you make it shorter.
[57:27] You can also build much denser switches
[57:29] and routers in a single rack up to
[57:31] realistically 6.4 petabits per rack.
[57:34] Today, that would be 1.6. And that
[57:36] means, you know, you because today
[57:38] you're limited by the front panel
[57:39] density of the OSFP.
[57:41] And finally, again, you know, nobody
[57:43] wants more racks. So, um
[57:46] one So, the questions people are asking
[57:49] is, you know, will it support a linear
[57:50] channel? That's like the number one
[57:52] question.
[57:53] The good news is, you know, we got our
[57:54] first modules literally last week. We
[57:57] plugged them in and the very, very first
[58:00] modules delivered 10 to the minus eight,
[58:02] you know, perfect beta rate, you know,
[58:05] straight out of the box. You know, it
[58:06] was amazing. And then the electrical
[58:09] channel, as it was tested here by a full
[58:11] retimed module, is at 10 to the minus
[58:13] 10. So, this is a linear, sorry, a
[58:16] flyover cable CPC kind of connector
[58:20] that creates a very good linear channel
[58:22] at 200 gig per lane. And these are just
[58:24] the first results. We're very happy with
[58:26] those results.
[58:28] Second, reliability.
[58:30] You know, so,
[58:31] we project the reliability to be far
[58:33] lower than conventional optics for four
[58:34] reasons. Number one, well, there's fewer
[58:37] lasers
[58:38] because, you know, you have more
[58:39] channels, but that's also true for
[58:40] conventional modules.
[58:42] You have much lower temperatures. You're
[58:44] coming in with a liquid fluid at 45
[58:46] degree. It's It's so the temperature
[58:49] environment is typically 20, 25% below
[58:51] air cooled.
[58:52] You have very little temperature
[58:53] variation, much less thermal stress on
[58:56] the modules.
[58:57] You have no vibration, mechanical
[58:59] vibration from fans in a liquid cooled
[59:01] system. All of this improves
[59:03] reliability. Then there's much fewer
[59:05] components. In the XPO, there's two
[59:07] paddle cards. Each one is 32 channels.
[59:10] And you need 75% fewer common components
[59:12] like microcontrollers and voltage
[59:14] converters compared to OSFP. And
[59:16] obviously, the most reliable component
[59:18] is the ones you don't have. And finally,
[59:20] you know, internal lasers that are
[59:22] directly connected are actually much
[59:24] more more reliable than external ELS
[59:27] based on vendor data, we we project
[59:29] these individual lasers at 45 to be
[59:32] below 1 fit. So, if you have eight
[59:34] lasers in the module and you throw in
[59:36] the linear drivers, it's a 20-fit
[59:39] addition compared to a equivalent laser
[59:42] module without the lasers. So, it's
[59:44] actually the lasers are not the problem
[59:45] here.
[59:46] So, let me talk about, you know, what
[59:48] surprised us the most what in the last,
[59:50] say, 6 months or actually this started
[59:53] at less than a year ago.
[59:55] So, the number one was actually the
[59:56] improvement in reliability which was
[59:58] actually larger than we initially
[59:59] estimated. It's an order of magnitude
[01:00:01] better than conventional modules.
[01:00:04] Now, the other big improve uh
[01:00:07] surprise was the tremendous amount of
[01:00:09] interest we're getting for
[01:00:11] embedding coherent light and full CR in
[01:00:14] this very small form factor. It turns
[01:00:16] out um the reliability improvement is
[01:00:19] even more relevant for coherent because
[01:00:20] you have so many more components. But,
[01:00:22] based on what our partners are telling
[01:00:25] us, you know, putting 12.8T terabits of
[01:00:27] coherent into a single module with fixed
[01:00:30] lasers which are more reliable than
[01:00:31] tunable is viable. You know, and you
[01:00:33] will see some of these products shown as
[01:00:35] as small prototypes at this show.
[01:00:38] Then, third, there's actually a lot of
[01:00:40] people that are working on 16 channel
[01:00:42] and 32 channel silicon photonics, a
[01:00:44] surprising number.
[01:00:45] In an OSFP and eight-channel OSFP
[01:00:48] module, you couldn't put them in because
[01:00:49] you're limited by the eight-channel
[01:00:51] density. But, they have a natural home
[01:00:53] in these higher channel count uh paddle
[01:00:55] cards. Then, the interest from uh
[01:00:58] microwave and copper is almost there.
[01:01:01] And by copper, we mean re-timed copper.
[01:01:03] So, what you do is you put the re-timers
[01:01:05] or the re-drivers inside the module and
[01:01:08] you can actually plug in the passive
[01:01:09] copper cable from that module to another
[01:01:12] module, which is pretty cool.
[01:01:13] And finally, perhaps the most surprising
[01:01:15] thing is the incredible interest we got
[01:01:18] on the XPMSE
[01:01:20] which was formally announced just a few
[01:01:22] days ago.
[01:01:23] We now have 60 60 members including 20
[01:01:27] of the world's leading optics module
[01:01:29] vendors.
[01:01:30] And with the permission, I can actually
[01:01:32] show you one slide of the logos.
[01:01:35] So, basically everybody that makes um
[01:01:39] optics modules in volume is on this
[01:01:40] list. There's also a few
[01:01:42] smaller companies or new companies on
[01:01:44] the list you may have never heard of
[01:01:45] before. But we have basically a lot of
[01:01:47] people joining us that want to make
[01:01:49] these kind of modules.
[01:01:51] So, to sum this all up, you know, we we
[01:01:53] think of this as a a form factor that
[01:01:55] solves
[01:01:56] short-term pain points including
[01:01:59] density, liquid cooling, high
[01:02:00] reliability, lower power, lowest power.
[01:02:03] Reducing structural costs is surely a
[01:02:06] cost-benefit, right? Reducing data
[01:02:08] center footprint is a major benefit.
[01:02:11] Being able to support all the different
[01:02:13] optics standards from copper, RF
[01:02:15] microwave to you know,
[01:02:18] coherent light to CR is great.
[01:02:21] Uh and finally,
[01:02:23] and and this is I think is a point
[01:02:24] that's underappreciated.
[01:02:26] Pluggable optics enable optics
[01:02:28] innovation cuz you don't have to decide
[01:02:31] today what you're going to bundle with
[01:02:32] your future CPU switch in 2 years from
[01:02:35] now, right? So, because the schedules
[01:02:37] are decoupled, you can actually plug in
[01:02:40] the thing whatever works the best in the
[01:02:42] future. And this is an inherent
[01:02:43] advantage.
[01:02:45] So, my last slide is, you know, we're
[01:02:47] here in Hollywood, LA. They would call
[01:02:50] it it's a wrap. I guess I'm also the
[01:02:51] last speaker for this session. The
[01:02:53] movie's in the can and we are ready to
[01:02:55] go to post-production with XPO.
[01:02:59] Thank you very much.
[01:03:05] I It is indeed hard to see, but while
[01:03:07] we're getting the audience warmed up for
[01:03:09] questions, I'll take the opportunity and
[01:03:11] I'll put Andy on the spot for a moment.
[01:03:15] Um how come in your uh table with nine
[01:03:18] technology options, there were no micro
[01:03:20] LEDs? Was this you ran out of space or
[01:03:23] what happened?
[01:03:24] I do believe your company actually
[01:03:26] joined the consortium, but I'm not sure.
[01:03:28] So, yes, you can build a micro LED XPO,
[01:03:31] absolutely. Okay, thank you. Just wanted
[01:03:34] to clarify. Let's see, where do we have
[01:03:36] questions in the audience? I have a few
[01:03:39] myself. Oh, I see you far over there.
[01:03:42] Okay, yeah, please. Please right, I
[01:03:44] think.
[01:03:44] >> Brian here from Terahop. Uh this is a
[01:03:46] question for you.
[01:03:48] Uh actually follow up earlier comment uh
[01:03:51] from Hong
[01:03:52] uh at Google.
[01:03:54] Uh literally people are building silicon
[01:03:56] photonics inside the transceiver, so you
[01:03:58] can actually And he also mentioned there
[01:04:00] were
[01:04:01] 16-channel PIC, 32-channel PIC you can
[01:04:04] build inside transceivers. To some
[01:04:06] extent,
[01:04:07] you can now build CPO inside the
[01:04:10] transceiver.
[01:04:11] Can you make a comment when you stating
[01:04:14] CPO is more reliable than pluggable by
[01:04:17] 5X. How do you compare those? Are you
[01:04:20] really comparing apple to apple in that
[01:04:22] case?
[01:04:23] Please comment if I do bring the very
[01:04:27] integrated
[01:04:28] PICs inside the transceiver. Are you
[01:04:31] going to see that 5X advantage?
[01:04:35] So, I I think it's unfair, right, for
[01:04:38] you to ask me to predict what the future
[01:04:40] looks like, right, for reliability when
[01:04:43] we're spending all this time and effort
[01:04:44] to try and pull forward the sort of data
[01:04:46] that we have. I can say with confidence,
[01:04:49] right, that that we believe in our data.
[01:04:51] I will update it and say that if you
[01:04:53] look at all the failures for the
[01:04:54] pluggables versus all the failures for
[01:04:56] the CPOs 2X, right? When when I pulled
[01:04:59] out the issues associated with the
[01:05:01] surface-mount component on the PLS, the
[01:05:05] ELSFP, right, it became 10X. So, I think
[01:05:08] those numbers are some X in there. If
[01:05:11] you're saying you're suggesting to me
[01:05:12] that XPO will be more reliable than a
[01:05:16] eight-channel pluggable module, I mean,
[01:05:18] I think I'll go back to Andy's like
[01:05:20] aphorism, right? Fewer parts will be
[01:05:22] more reliable than more parts. So, I we
[01:05:24] would expect that a pluggable XPO will
[01:05:26] be more reliable
[01:05:27] on a per port basis than eight
[01:05:30] individual eight-channel devices.
[01:05:33] However,
[01:05:34] I don't know that it's going to get to
[01:05:35] the same as CPO.
[01:05:37] Time will tell.
[01:05:39] Yeah, I think one of the key part here
[01:05:41] is you compare practically old
[01:05:43] technology built on a OSP you are
[01:05:46] testing today
[01:05:48] versus a CPO not yet to be deployed. So,
[01:05:50] I think only fair to see
[01:05:52] what is available
[01:05:54] in terms of silicon photonic base
[01:05:56] transceiver versus
[01:05:58] a CPO. Also, taking the DSP out because
[01:06:01] some of these failure is coming from
[01:06:03] DSP. So, I think you're testing LPO as
[01:06:05] well. I think it will be good really
[01:06:07] great test to compare a silicon photonic
[01:06:10] LPO
[01:06:11] versus a CPO because that's sort of fair
[01:06:14] comparison.
[01:06:15] >> So,
[01:06:16] I agree with you. However, I will say
[01:06:18] that when we look at pluggable modules,
[01:06:20] the dominant failure mode has nothing to
[01:06:22] do with silicon photonics or EMLs. So, I
[01:06:24] would challenge you that the issue is
[01:06:26] silicon photonics versus non-silicon
[01:06:29] photonics.
[01:06:30] If if I'll I'll I'll maybe we take it
[01:06:33] offline. We have this debate around
[01:06:34] pluggables and if they're more or less
[01:06:36] reliable, but I can only tell you what
[01:06:38] the data shows for the systems that we
[01:06:39] have. If I could add to this, so
[01:06:43] we are actually planning to do an
[01:06:44] extensive life cycle test with XPO later
[01:06:47] this year. That's equivalent to the test
[01:06:50] that, you know, Meta has been running
[01:06:51] now for some time. So, we'll have, you
[01:06:53] know, hundreds of switches fill up with
[01:06:55] XPO getting millions of devices to truly
[01:06:58] demonstrate both soft and hard error
[01:07:00] rates. And we don't have the data yet,
[01:07:02] so I can't predict the future, but what
[01:07:04] we expect is that if the XPO module
[01:07:07] built with the same components as CPU or
[01:07:10] the equivalent technologies and some of
[01:07:12] components, same temperature, same
[01:07:15] environment and so on, we expect to see
[01:07:17] very similar failure rates. Again, we
[01:07:19] don't have the data yet. I hope to
[01:07:21] report on this next year or or even
[01:07:23] later this year, but we will run that
[01:07:25] test on behalf of the industry.
[01:07:30] Any other audience questions?
[01:07:33] If not, I have one more question for the
[01:07:35] for the panel, for for everyone. We we
[01:07:38] we sort of looking at at four attributes
[01:07:40] now that came up time and again. It's
[01:07:42] reliability and cost. If we don't meet
[01:07:44] those with new interconnect
[01:07:46] technologies, it's a no-go.
[01:07:48] And then power and reach. Power is a
[01:07:51] nice to have because it's always better
[01:07:52] to have lower power, but of course reach
[01:07:54] is probably more important, but I would
[01:07:56] like to hear from the from the panelists
[01:07:58] on how they think about these four these
[01:08:00] four components. If I could comment on
[01:08:03] power really quick. So, in a linear
[01:08:05] channel environment, right? Linear to
[01:08:07] linear,
[01:08:08] um the XPU is actually basically the
[01:08:11] same power as CPU.
[01:08:14] You're you have less optical coupling
[01:08:16] losses, so the laser power is actually
[01:08:17] lower.
[01:08:19] One could argue that you need slightly
[01:08:20] higher drive current and maybe the
[01:08:21] electrical power is slightly higher, but
[01:08:23] it's either a wash or lower than CPU.
[01:08:26] So, maybe I'm a comment. I think the
[01:08:30] the primary driver is going to be
[01:08:31] somewhat temporal, right? Right now,
[01:08:32] what we're seeing is uh what is driving
[01:08:35] a need for deeper levels of optical
[01:08:37] integration into scale-up systems is
[01:08:40] performance. And the performance, by the
[01:08:41] way, that one of the metrics you're
[01:08:42] missing is latency, all right? So,
[01:08:44] latency translates directly to revenue
[01:08:45] for a lot of these applications. So, it
[01:08:47] is uh it's bandwidth and latency and
[01:08:50] interconnectivity, right? As the as the
[01:08:52] as the scale-up cluster domains
[01:08:54] increase, size domains increase.
[01:08:57] I believe and and and we believe that
[01:08:59] over the course of next 2 3 years that
[01:09:02] power, because the limitations in energy
[01:09:04] availability in the United States,
[01:09:06] right? The the demand for data center
[01:09:08] power consumption is going to double the
[01:09:09] next 4 years.
[01:09:10] Um so that is the limiter for the
[01:09:12] ability to stand up incremental data
[01:09:14] center capacity domestically and in most
[01:09:18] places throughout the world and that
[01:09:19] will increasingly become
[01:09:21] a contributing factor to all
[01:09:23] infrastructure.
[01:09:24] I can also agree with Andy's input that
[01:09:26] LPOs Can Can I clarify really quick? So
[01:09:28] I was talking about high-speed CPU. Slow
[01:09:31] and wide, you know, which is I think
[01:09:32] what you're talking about, is in fact
[01:09:34] lower power than high-speed because
[01:09:35] you're limiting the high-speed service,
[01:09:37] right? That is true.
[01:09:38] On the other hand, the risk factor is
[01:09:40] you're bundling a future technology to
[01:09:42] your next generation GPU, XPU, switch.
[01:09:45] If it doesn't ship on time, you can't
[01:09:47] ship. So the fundamental risk is just
[01:09:49] time to market and being able to scale
[01:09:51] this. And that's I think the the
[01:09:53] singular biggest issue with integration.
[01:09:56] Yeah,
[01:09:58] not to debate on the stage. We're
[01:09:59] relatively agnostic at this point, at
[01:10:00] least I am now that I'm a a Marvell
[01:10:02] employee instead of a Celestica CEO.
[01:10:05] We're we're agnostic to the
[01:10:06] implementation. So yeah, we're
[01:10:08] we're not uh religious about any
[01:10:10] implementation of optical. Okay, thank
[01:10:12] you. Peter, Drew, any Yeah, maybe maybe
[01:10:15] one comment on cost.
[01:10:17] So and that that amplifies Andy's points
[01:10:19] that he made before that
[01:10:21] coupling single mode fibers is not for
[01:10:23] come does not come for free and is not
[01:10:24] as scalable.
[01:10:26] If you if you look at the single mode
[01:10:28] optics for scale up um
[01:10:31] if you really want to replace copper,
[01:10:33] good luck from a cost point of view. It
[01:10:35] won't work because just the just the
[01:10:39] connection of the fiber each fiber
[01:10:41] connection costs you a dollar.
[01:10:43] So you have a dollar of the FAU and in
[01:10:45] high volume, a dollar of at the FAU, a
[01:10:48] dollar at the front panel, another
[01:10:49] dollar from the patch cord you plug in
[01:10:51] and then you go to the other side. So
[01:10:53] that's at least 3 cents a gigabit right
[01:10:56] there and you have not done any silicon
[01:10:59] photonics, you have not powered it, you
[01:11:00] have no laser, you have nothing. So,
[01:11:03] um
[01:11:04] competing with copper in the scale up is
[01:11:07] going to be very challenging on a cost
[01:11:09] perspective. So, that's why mixed media
[01:11:11] is super important.
[01:11:15] I guess it's a panel to be
[01:11:16] controversial. Um
[01:11:18] we believe optical scale up can be cost
[01:11:21] competitive with copper, full stop
[01:11:23] period.
[01:11:24] Have to prove it, much like Roy I
[01:11:25] believe for NPO, CPO, and pluggable
[01:11:28] modules, right? But but there is math we
[01:11:30] believe which gets us there.
[01:11:32] And that includes all the fibers and
[01:11:34] everything?
[01:11:34] >> everything soup to nuts, direct
[01:11:36] comparison of a copper backplane versus
[01:11:39] optical scale up. You're not alone. Like
[01:11:41] rack to rack, other cables and things
[01:11:43] like that add and change that and and
[01:11:44] there's another aspect to that, but but
[01:11:47] that that that that'll be a point of
[01:11:49] contention. I guess I want to touch on
[01:11:50] power though, too, and I think this is a
[01:11:52] really important point and goes back to
[01:11:54] the sort of infrastructure comments that
[01:11:56] Andy was making, but also Hong made
[01:11:58] earlier.
[01:12:00] Um
[01:12:01] you know, the power consumed in the data
[01:12:03] center by networking is not the dominant
[01:12:05] contribution of power, right? And it's
[01:12:07] great to save that power and and we
[01:12:09] should do that and that'll be great.
[01:12:11] From our perspective, it's GPU cost,
[01:12:13] right? If we can save
[01:12:15] you know, X amount of power and deploy X
[01:12:18] many more GPUs in the same data center
[01:12:20] in the same power envelope, that is
[01:12:22] nothing but win, right? That's an
[01:12:24] asymmetric sort of value proposition
[01:12:26] because the value of each of those GPU
[01:12:28] flops or number of tokens can create
[01:12:30] enormous amount of value. That's really
[01:12:32] the win, not the are we saving half a
[01:12:34] picojoule per bit here or not, but
[01:12:35] really what that means is to Andy's
[01:12:37] point, right? If we're going to do
[01:12:39] co-packaged optics, if we're going to
[01:12:40] bet on OCI, it has to come at the same
[01:12:42] time because you have to bet your
[01:12:44] infrastructure on it. You can't have a
[01:12:45] backup which is like twice as expensive
[01:12:48] with retimed optics and say, "Okay,
[01:12:50] well, the the most reliable thing is
[01:12:52] that I put in and make it sure that it
[01:12:53] works for re-timed optics, but I'll save
[01:12:55] the X amount of power in like 1% of the
[01:12:58] data center. It's the fact that I can
[01:12:59] take out that that that backup strategy
[01:13:02] because I believe it because it works
[01:13:04] because it it arrives on time, and then
[01:13:06] I can put the extra GPUs in. That's a
[01:13:09] big bet. It's a big statement, and it's
[01:13:10] hasn't been proven, but that's what
[01:13:12] we're driving to.
[01:13:13] Can I one one one more comment on power,
[01:13:15] which is
[01:13:17] um
[01:13:18] obviously the ultimate goal is to
[01:13:19] improve utilization of the GPUs, right?
[01:13:22] And it is true if you had more bandwidth
[01:13:25] per GPU, you get better utilization,
[01:13:27] which pays for all the power in the
[01:13:29] optics, right? So, the real goal is to
[01:13:31] get a lot more bandwidth. So, the number
[01:13:32] projected to go from
[01:13:34] uh 25.60 to 100 to 4 P per GPU is a
[01:13:38] clear goal because the the GPUs need
[01:13:41] much more bandwidth than we ever thought
[01:13:43] to improve the efficiency, right? Yes,
[01:13:46] you will have more optics and more
[01:13:47] optics power, but your GPU efficiency
[01:13:50] increases a lot more.
[01:13:53] Okay, let's go to Drew. You have a
[01:13:55] question? Hi, everyone on the panel
[01:13:57] knows me, of course, and probably most
[01:13:59] of the people in the audience do, too,
[01:14:00] but for those of you don't, I'm Drew
[01:14:01] Perkins, and I'm CEO of Airdew, but I'm
[01:14:03] also
[01:14:04] uh on the board of Mojo Vision. And so,
[01:14:06] Chris, you took uh Andy to task for not
[01:14:09] mentioning micro LED, but you didn't
[01:14:11] take uh Dave to task there because they
[01:14:13] have a deal with Mojo Vision, and just
[01:14:15] wanted to point that out.
[01:14:17] Well, thank thank you for
[01:14:19] >> Thanks, Drew. You're the best.
[01:14:21] Um
[01:14:22] did we have
[01:14:24] question over there?
[01:14:25] Yeah, hi, this question is for Andy.
[01:14:27] Andy, in your comment about um getting
[01:14:29] to a 1 billion equivalent of
[01:14:31] transceivers, and you in that same slide
[01:14:33] you said that, you know, that
[01:14:35] non-optical was would be the the the
[01:14:37] best the the
[01:14:39] a different way to get to that that same
[01:14:41] end point. So, I guess the question is
[01:14:42] is do you feel the the path of of least
[01:14:45] resistance is to get the market to be
[01:14:47] able to support a billion optical
[01:14:49] devices or to find innovation that
[01:14:52] allows you to have copper extend longer.
[01:14:55] And then
[01:14:56] the site really meant
[01:14:58] you know, today it's all copper.
[01:15:00] And everybody wants to get to much
[01:15:02] larger clusters like 512 1024 which is
[01:15:05] beyond the reach of copper, right?
[01:15:07] However, the physical dimensionality of
[01:15:09] 512 which is still within the reach of
[01:15:12] other low power technologies which are
[01:15:14] in fact very cheap and very reliable and
[01:15:17] very robust. So I think that is under
[01:15:19] appreciated. In other words, you can
[01:15:21] connect things very easily in the next
[01:15:24] few years in 10 or 20 meters which is
[01:15:26] with quality alternative copper
[01:15:28] technologies.
[01:15:30] But that is the most real competition to
[01:15:32] future single mode.
[01:15:36] Any more questions from the audience?
[01:15:39] We got two more minutes. Um
[01:15:42] okay, while you think of another
[01:15:43] question I'll I'll make another comment
[01:15:45] because cost came up and in my
[01:15:48] experience I've been in this industry
[01:15:49] for about 30 years. I don't think there
[01:15:51] are any companies better than the large
[01:15:53] cloud service providers to beat up on
[01:15:56] the optical community and get their
[01:15:58] price to where it needs to be to be
[01:16:00] competitive. They have the mass, they
[01:16:02] have the the skill to do that and I
[01:16:05] think they will deliver again. Just just
[01:16:07] watch.
[01:16:08] We got one more question.
[01:16:11] Yeah, over here.
[01:16:11] >> Okay, please.
[01:16:13] Hi, this is Bill from Credo. Just to add
[01:16:14] to the controversy, there was a panel
[01:16:16] earlier that they were asked to put a
[01:16:19] stake in the ground about 2028
[01:16:21] and what percentage will be CPO?
[01:16:24] That was for scale out. I'd like to hear
[01:16:26] your opinion for scale out and scale up.
[01:16:30] Starting with Andy. Well, I mean Nvidia
[01:16:34] obviously is really betting on CPO,
[01:16:35] right? So if there's one company that's
[01:16:37] putting the metal to the pedal or
[01:16:39] whatever you say, it's Nvidia and they
[01:16:41] they have such a market share that they
[01:16:43] could ship substantial numbers of
[01:16:46] CPU once it's in full production, right?
[01:16:48] That's no question about that. It's the
[01:16:51] we I was talking about the rest of the
[01:16:52] industry X Nvidia, right? And then you
[01:16:54] look at, you know, what it really takes
[01:16:57] to follow Nvidia's lead to make that
[01:16:59] into high volume manufacture product, it
[01:17:01] takes a lot of effort.
[01:17:05] Yeah, Bill. In addition to Nvidia,
[01:17:08] right, pushing a large part of the the
[01:17:10] shift. We're talking 2028 totals
[01:17:14] scale out scale up. So, Jensen just
[01:17:16] announced
[01:17:17] the the scale up strategy for NVLink 8.
[01:17:21] We'll include some
[01:17:23] some CPO and then they've got the
[01:17:24] Spectrum-X strategy. So, that that'll be
[01:17:26] a big chunk that pushes things toward
[01:17:28] the higher end.
[01:17:30] Broadcom, obviously, supporting that
[01:17:32] direction. We're strong believers that
[01:17:35] incremental adoption in this period as
[01:17:36] well. A few hyperscalers are very very
[01:17:39] aggressive with their shift to
[01:17:42] CPO-based scale up connectivity. So, I
[01:17:45] actually agree with Near's comment
[01:17:47] earlier, closer to about 30% by by 2028.
[01:17:53] I think also that we'll have at least
[01:17:55] 30% by 2028, but I'm not an analyst and
[01:17:58] I said that on the Signal AI webinar
[01:18:00] already when I was asked the same
[01:18:01] question.
[01:18:04] Drew, anything to add?
[01:18:05] Nope. Okay. Otherwise, it's a it's a
[01:18:08] wrap. We're out of time. Thank you
[01:18:09] again. Thank you to the speakers.
[01:18:26] Everybody talking
[01:18:28] about scaling AI
[01:18:32] but the data center's joking deep
[01:18:34] within.
[01:18:38] Copper running hot.
[01:18:41] Yeah, the signal's getting thin.
[01:18:44] So, we flip the switch now.
[01:18:48] Optics is in.
[01:18:50] Bandwidth climbing fast.
[01:18:53] Racks are running red.
[01:18:56] Cloud demand exploding overhead.
[01:19:02] Pluggables fading as the limit's
[01:19:06] closing.
[01:19:09] Co-packaged light is how we win.
[01:19:14] It's photonics,
[01:19:16] baby. It's 2026.
[01:19:20] Riding that light wave.
[01:19:23] Doing new tricks.
[01:19:25] From the fiber in the ground to the chip
[01:19:28] in my hand, we make that sunshine jump
[01:19:31] on command.
[01:19:32] Yeah, photonics, baby.
[01:19:37] 2026.