# Scaling AI Networks: Incremental Innovation or Photonic Disruption?

https://www.youtube.com/watch?v=YBjyzucZJ3c

[00:01] So, hello everyone.
[00:03] This is the last uh session before lunch.
[00:06] Uh hopefully it'll be informative and fun.
[00:09] Um we're going to start very briefly with each company giving having a single slide just to try and explain in a in what's an unfair way because it's such a short window, but to try and give you a sense of what each company is doing.
[00:24] And then we'll go into the question and well, we'll we'll start asking questions and at some stage I'll open it up also to to the audience.
[00:32] So, to start with, can I can I please have Jeremy from Astera Labs?
[00:38] And just to talk through the slide, but if you if you're happy doing it there, that's even better.
[00:45] Good morning.
[00:47] Um Next slide, please.
[00:50] One more.
[00:51] Thanks.
[00:53] Yeah, so it's a it's a pleasure to be here this morning.
[00:56] So, Astera Labs is a company that provides a complete solution portfolio
[01:01] provides a complete solution portfolio for AI rack-scale connectivity.
[01:04] for AI rack-scale connectivity.
[01:06] So, special emphasis on very robust solutions, low power, scalable solutions.
[01:09] solutions. And goes all the way from uh sing- signal conditioners, uh memory controllers,
[01:15] um electrical uh switch fabrics both uh for the front end of the network.
[01:21] And all the the latest high radix products uh for the uh scale-up network.
[01:27] And uh optical and electrical uh interconnects.
[01:30] Right. So, our latest focus is to um optically enable scale-up networks for multi-rack uh compute nodes.
[01:42] So, company was fund- founded in 2017.
[01:45] Uh it's been the fastest company in the semiconductor segment to reach a billion US dollars in in revenue.
[01:54] Um and yeah, this this slide here shows um the the the products that have been introduced over time that has been going to increasing level of complexity
[02:04] to increasing level of complexity to basically cover everything that's to basically cover everything that's connectivity inside of connectivity inside of electric.
[02:07] Thank you.
[02:09] Next company is Credo and Chris Collins, please.
[02:15] Good morning.
[02:17] I think I'd like to start here.
[02:19] In AI infrastructure, for a lot of folks who maybe aren't at this conference, they're overly obsessed with the compute and not obsessed enough with the connectivity.
[02:28] And you know, if your interconnect is burning too much power or it's not reliable, it's it's dropping links, you've really just built a really expensive heater as your GPUs sit idle.
[02:40] So, we believe that um going forward, the real the solution is not optics or copper, it's yeah, deploying a the right mix of both with reliability and power efficiency built in.
[02:53] So, Credo is a pure play connectivity company.
[02:55] We're unique because we own the whole stack from the SerDes to the chip to the hardware design.
[03:01] And you can see from our portfolio up here, you know, we really pioneered the
[03:06] here, you know, we really pioneered the active electrical cable market for low.
[03:08] active electrical cable market for low power, high reliability, short reach links.
[03:11] We've recently introduced the active LED cable which extends that low power reliable approach to the end of the row.
[03:20] And then recently launched a zero flap transceivers which extend that high reliability out even further.
[03:27] So, what matters most I think for AI networks is not just pure bandwidth, it's bandwidth per watt, it's time to first token.
[03:35] And that's why we believe a single approach isn't the best, but a this scalable offerings with the focus on power efficiency and high reliability.
[03:47] Thank you.
[03:50] So, that's the first two established companies.
[03:53] Now moving on to two exciting startups.
[03:57] So, uh if I can ask ask Professor Calabretta to talk about Astra Networks.
[04:01] Yeah, thank you very much for the invitation.
[04:04] Very exciting to be here.
[04:06] And yeah, I'm Nicola Calabretta, I'm professor at Technical University Eindhoven and also co-founder of Astra Networks.
[04:13] In Astra, what we do is to introduce uh advanced photonics and electronics in order to deal with uh AI traffic and in particular the possibility to have this kind of tunneling and operate the traffic in a much more dynamic way.
[04:32] Now, the first things we are doing is Infinity 48, that is the first fabric that they allows indeed to have a physical network layer intelligent and programmable as a day software layer and possibility to interconnect 110,000 nodes with just zero to two retransmitting hops.
[04:53] The other advantage of what we introduce is the design that today also that is available for tomorrow.
[05:00] What we mean is that to have this kind of both predictable AI training for today but also a kind of unpredictable AI interface and deal with
[05:08] unpredictable AI interface and deal with the traffic as the volume increases.
[05:11] the traffic as the volume increases.
[05:14] Uh we also consider the nodes not just like switch nodes but also to have enabled in network computing and then makes also transport from every hop itself more intelligent.
[05:16] like switch nodes but also to have enabled in network computing and then makes also transport from every hop itself more intelligent.
[05:19] enabled in network computing and then makes also transport from every hop itself more intelligent.
[05:20] makes also transport from every hop itself more intelligent.
[05:23] itself more intelligent. And uh last but not least reduce the tail latency and that also improve the a lot the utilization of the resources.
[05:26] not least reduce the tail latency and that also improve the a lot the utilization of the resources.
[05:28] that also improve the a lot the utilization of the resources.
[05:30] utilization of the resources. Uh next products we are working on is 256 to scale to have further scaling.
[05:33] resources. Uh next products we are working on is 256 to scale to have further scaling.
[05:36] working on is 256 to scale to have further scaling.
[05:37] further scaling.
[05:39] And uh the main idea is indeed is a seamless integration of exciting architecture and standards protocols as well.
[05:41] seamless integration of exciting architecture and standards protocols as well.
[05:43] architecture and standards protocols as well. So, we are not going to change any kind of protocols but yeah, the network is still operating with legacy protocol.
[05:45] well. So, we are not going to change any kind of protocols but yeah, the network is still operating with legacy protocol.
[05:48] kind of protocols but yeah, the network is still operating with legacy protocol.
[05:51] is still operating with legacy protocol. So so just to add you're in stealth mode at the moment, but just to understand is it right that from the conversations that we had earlier today explaining the hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[05:53] So so just to add you're in stealth mode at the moment, but just to understand is it right that from the conversations that we had earlier today explaining the hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[05:55] at the moment, but just to understand is it right that from the conversations that we had earlier today explaining the hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[05:57] it right that from the conversations that we had earlier today explaining the hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[05:59] that we had earlier today explaining the hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[06:02] hierarchy of switching, your your networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[06:04] networking is a combination of of electrical switching and some optical switching. Is that fair summary? Yeah,
[06:06] electrical switching and some optical switching. Is that fair summary? Yeah,
[06:08] switching. Is that fair summary?
[06:10] Yeah, we are uh developing we are uh developing kind of smart network interface that is kind of smart network interface that is the one that actually enable the fact to use electronics and photonics switches.
[06:18] and that allows the operation of in case like east-west traffic in a very very quick way and for high capacity as well.
[06:29] as a possibility of in network computing.
[06:32] Okay, so just just as the way that the company is seated, um the last speaker now is um um Paul from Oriole and you're an example of a startup that's actually looking to get rid of electrical switching.
[06:46] Is that Is that fair?
[06:49] That's right. We we're doing something unusual.
[06:50] So I'm Paul Wallace from Oriole Networks.
[06:54] We're based in London, but we've got offices around the UK, in the US and Europe.
[07:01] But one of the things that we believe is that if you if you're trying to build a new type of photonic network which has got the full realize the full potential of photonic networks, you need to kind
[07:09] of photonic networks, you need to kind of start again and rethink it.
[07:11] And I think the the interesting challenge there is is a lot of photonic companies are trying to build individual components and try to accelerate them.
[07:21] Whereas we believe to you've got to unlock that to do a lot of it.
[07:24] So very similar to to my friend here, we're trying to rethink the way that works.
[07:28] So we're actually we've done series A, we're quite quite new, but we're now opening for a series B as well.
[07:33] And I think as we we'll look at this, a lot of the innovations come from university type scenarios, but we're mixing that with high volume manufacture and traditional traditional design to get something which more people can can use and and access.
[07:48] Okay.
[07:50] So, moving on now to the the Q&A part, if I can ask each of you very very briefly, but to put it in your own words, can you explain what problem you're solving as part of this whole world of AI-based data centers?
[08:07] Okay, I'll start with that.
[08:09] Okay, I'll start with that.
[08:11] So, if you look at uh.
[08:12] So, if you look at uh Yuval Barenshtein this morning from Nvidia said the number one problems for networks are jitter and congestion and that sort of thing.
[08:19] They're the ones that are the enemies of workloads.
[08:20] But, if you're trying to drive more and more data through electrical optical conversions, you're going to create more and more of those challenges.
[08:29] Yes, and more of those challenges.
[08:30] Yes, there's techniques and and and so on.
[08:33] But, as you start We're seeing as you start to go to 1.6, 3.2 terabits, it really really starts to get very very hard.
[08:38] And so, the approach I think is to to to change that and think differently.
[08:40] Design a network which is uncongested, can deterministic latency, and move on from that.
[08:45] So, it's about performance.
[08:47] It's about removing the challenges of the electrical network by different.
[08:53] Yep.
[08:54] Would you agree with that, Professor?
[08:55] Would you add anything?
[08:56] >> Yeah, indeed.
[08:58] That's That's actually the number one of the must to do, right?
[09:00] Is to improve the congestion in a way that in such way we have a better utilization of the network.
[09:04] And obviously also to introduce this where is needed the photonics in order to actually to remove
[09:11] Photonics in order to actually to remove a lot of power consumption is there.
[09:14] A lot of power consumption is there.
[09:17] Still the requirements is also the fact that how we deal with not only layer low layer layer one, but also how to deal with the other layers.
[09:24] That's very important ultimately because the services will be will be requires all the layer to be taken care.
[09:31] And that's actually where our companies also include a lot of networking people in order to have combination between the lower layer and the photonics that's very advanced together with also higher layer.
[09:46] Any more any either of you want to give your own perspective of with the problems that you're solving?
[09:54] I mean at a high level, we I think we all share the same mission, right?
[09:57] So, make sure that that's um the the communication does not bog down the compute.
[10:03] Um, but, Astera Labs specifically aims at being a one-stop-shop place where you can get all the components you need for that and to have, you know, with the
[10:11] that and to have, you know, with the unified telemetry system uh.
[10:14] unified telemetry system uh that orchestrates very efficiently all those devices.
[10:17] those devices. Uh looking ahead,
[10:19] one of the the very important missions is to make sure that the the multi-rack scale-up network that will be hybrid electrical optical.
[10:26] scale-up network that will be hybrid electrical optical inherits the same efficiency in paradigms than the all electrical networks we have today. Okay.
[10:35] electrical networks we have today. Okay. >> very important thing we're working on.
[10:40] Yeah, I guess I would echo a lot of the same sentiment.
[10:43] You know, Credo is really um addressing data movement within AI infrastructure uh with a heavy focus on reliability and using the right tool for the right job, you know, so for shorter reaches um whether it's an active electrical cable, micro LED, and then moving up to the zero flop uh transceivers for the longer reaches.
[11:06] longer reaches. Okay, another question and it may work and it may not, but do you Do any of you have a telling metric or a telling
[11:12] have a telling metric or a telling anecdote with your interactions that shows the urgency of AI?
[11:15] anecdote with your interactions that shows the urgency of AI?
[11:18] Does anything come to mind that when you first heard it I I do have one.
[11:19] come to mind that when you first heard it I I do have one.
[11:22] So, we were working with a large customer and you know, if you consider a few years ago a big cluster maybe it was tens of thousands of GPUs and now we're moving upwards of a million GPUs.
[11:24] So, we were working with a large customer and
[11:27] customer and you know, if you consider a few years ago a big cluster maybe it was tens of thousands of GPUs and now we're moving upwards of a million GPUs.
[11:28] you know, if you consider a few years ago a big cluster maybe it was tens of thousands of GPUs and now we're moving upwards of a million GPUs.
[11:30] ago a big cluster maybe it was tens of thousands of GPUs and now we're moving upwards of a million GPUs.
[11:32] thousands of GPUs and now we're moving upwards of a million GPUs.
[11:35] upwards of a million GPUs.
[11:38] So, problem that they were having that maybe showed up, you know, once a day or something like that a few years ago when you move to a scale of a million GPUs shows up every minute or every 30 seconds.
[11:41] maybe showed up, you know, once a day or something like that a few years ago when you move to a scale of a million GPUs shows up every minute or every 30 seconds.
[11:43] something like that a few years ago when you move to a scale of a million GPUs shows up every minute or every 30 seconds.
[11:45] you move to a scale of a million GPUs shows up every minute or every 30 seconds.
[11:48] shows up every minute or every 30 seconds.
[11:49] seconds.
[11:50] Um and so improving things like the reliability of that connectivity was a big deal for them.
[11:52] and so improving things like the reliability of that connectivity was a big deal for them.
[11:54] reliability of that connectivity was a big deal for them.
[11:58] So, so I would say the the the the capacity that we're being asked to create for some of our products is is really testament uh to how much urgency there are to things.
[11:59] So, so I would say the the the the capacity that we're being asked to create for some of our products is is really testament uh to how much urgency there are to things.
[12:01] the the the capacity that we're being asked to create for some of our products is is really testament uh to how much urgency there are to things.
[12:04] asked to create for some of our products is is really testament uh to how much urgency there are to things.
[12:06] is is really testament uh to how much urgency there are to things.
[12:07] uh to how much urgency there are to things.
[12:10] to how much urgency there are to things.
[12:12] Um unfortunately it doesn't befall to me to give details. So, but it's uh
[12:15] To give details.
[12:15] So, but it's uh it's something we live every day.
[12:19] Of course we are not established company like you have.
[12:22] So, you have a much more anecdotes probably to tell.
[12:24] But one of our main point to talk with customers is about how to enable indeed the introduction of this photonics to save a lot of energy and who is capable to control them in a good way.
[12:36] And so far all the technique used actually they mostly failed it or especially when you bring those device much closer to the processing part.
[12:47] And that's one of the big issue that actually has prevented a lot so far even my small but large history on the photonic guys that prevent those device really to be much more in the market.
[13:01] Okay.
[13:02] Yeah, I think I can echo that.
[13:04] The the energy um imperative a lot of people are there is is really attracting and so on.
[13:09] And and from our perspective we have this idea of we have a completely passive glass core.
[13:14] So, we don't want the
[13:16] glass core.
[13:16] So, we don't want the intelligence around the edge.
[13:17] And so it intelligence around the edge.
[13:17] And so it sounds as though with zero zero power in the network that sounds attractive.
[13:22] What really came across to us is that what people are now looking at is actually it's the time to complete.
[13:28] Uh uh uh time to complete work workloads, time to complete steps, MPI, that sort of thing, collectives.
[13:36] Because actually, once you take take away the congestion in the network, it turns out that actually, you've got low jitter, low congestion, lower latency, they're deterministic.
[13:45] And that people are now attracted to the performance of of photonics, and you're probably saying the same, right?
[13:50] Where actually, that that then leads to a real business value in doing it beyond simply energy saving.
[13:58] So, if I can just ask start with the two startups, um what is the one thing that you would say has allowed your company to come into existence based on market dynamics?
[14:12] What basically led to your company being formed?
[14:14] I I think the the confluence of a number
[14:17] I I think the the confluence of a number of specific technologies that uh came.
[14:20] of specific technologies that uh came from University of London around fast laser tuning, burst rate recovery, and so on.
[14:25] But actually, it's the the ability to manufacture in low-power silicon photonics.
[14:29] That that's the enabler because we can now start to put that into small packages and make it useful and make it manufacturable because actually, volume will set you free in this sort of market.
[14:38] You need to be able to build reliably and and with resilience.
[14:45] Yeah.
[14:46] Yeah.
[14:46] Uh in um our case, the way how we started this adventure is is just the the explosion of AI in a way, that's mostly, and then the requirements how we actually to come with a paradigm to reduce power, reduce latency, deterministic, but also combine the networks also for inference, and have a possibility to have a system that develop deploy this.
[15:05] Okay.
[15:09] And just in to ask um these two established companies, but to ask it slightly differently, and I know things have been happening very recently, how do you make
[15:17] happening very recently, how do you make sure you've got the right technologies?
[15:19] sure you've got the right technologies going forward?
[15:21] going forward um to expand your IP to make sure that you're covered as things evolve.
[15:29] Well, um, so Achronix is very established in copper connectivity.
[15:35] Um, so we really see, uh, a big need now to to have optics not just in the scale out network, but really in the scale up, too.
[15:43] Um, and this is something we we're working very hard to to intercept, right?
[15:47] So make sure when, uh, CPU goes to the scale you know, that we're there.
[15:53] Um, yeah, that's that's that's >> So watch this space, yeah.
[15:57] Watch this space. Is that what you're saying?
[16:00] >> Yes, yes. Basically, look at the, uh, the changes in the, uh, the big the changes in the infrastructure that are being prepared and and be very early in that game.
[16:07] But just to just to add one detail, I take it you see still quite a bit of demand for copper and pushing copper.
[16:13] That's, um, can you say anything about that?
[16:19] um, can you say anything about that?
[16:21] Yes, so we I mean, I'm I'm an optics guy, right?
[16:22] And all my my career in optics.
[16:23] Uh, so so I'm I'm not the right person to ask, but as a company policy, uh, Achronix is very much covering both, right?
[16:32] And Okay. copper is not disappearing anytime soon.
[16:34] So And just to ask you to to comment because you recently made an acquisition, so if you can just very briefly comment on the the motivation.
[16:41] So I'm part of the acquisition.
[16:43] Um, I'm I'm one of the two founders of X-Cal.
[16:45] Uh, so I joined in in November.
[16:49] Uh, we are company We were a company that was doing, uh, detachable uh, fiber-to-chip, uh, connectors.
[16:58] Right. Um, so that's we we're going also as Achronix, we're going to proliferate that technology on the market.
[17:03] And of course, it's also an enabler for our own road map uh, for CPU.
[17:09] Okay. So, I guess that's as much as I can say.
[17:11] that's okay. And and Chris, you recently your company recently made an acquisition.
[17:15] We did. Actually a couple if you look back over
[17:20] Actually a couple if you look back over a few months.
[17:22] So, yeah, one way that we're preparing for you know, sort of next generation and making sure we have the tools.
[17:29] Obviously, there's the the service development, which is a core of what Credo does.
[17:32] The 200 G per lane today and then looking forward to 400 G per lane.
[17:40] But adding to that capability, the acquisition of Hyperloop for the micro LED part of our portfolio and then more recently DustPhotonics expanding on that the silicon photonics and making sure we have all of those key building blocks in house will be critical for us going forward.
[17:56] Okay, so before I shift gears, is there anyone who wants to ask a question based on what's what we just one question and then I'll move on.
[18:04] Any questions based on what you've heard so far?
[18:08] Okay.
[18:09] So, um next question, please.
[18:16] Um the um Sorry, I've I've got my papers jumbled.
[18:23] Yeah.
[18:25] Yeah.
[18:27] If you if you There's two types of companies here.
[18:29] So, There's two types of companies here.
[18:30] So, again, the startups you're bringing product to market, but you've also got your established road maps and serving existing customers.
[18:37] existing customers.
[18:37] But can you you talk each of you talk about some key intercept points coming up that you're aiming to get to?
[18:44] Whether it's a based on a particular transceiver speed or a certain type of switch fabric or however you answer it.
[18:51] But what are you trying to intercept?
[18:52] What's on the horizon?
[18:54] Yeah, I think in the near future the the big intercept is is the 1.6 T market.
[19:00] Of course, for active electrical cables, the Credo zero flat transceivers.
[19:05] We want to make sure that we're prepared for those in the audience who don't understand zero flat, can you just very briefly explain what what when you keep saying that?
[19:12] Yeah, so those transceivers that we've introduced, you know, it's a new class of transceiver with built-in telemetry and intelligence into the DSP that goes inside of those that enables the the
[19:25] inside of those that enables the the user to extract more information from your link.
[19:30] So, you can predict before a failure or a link flat might happen, take predictive or preventative maintenance.
[19:35] Um and then in the the other piece of the puzzle is actually having communication between the two transceivers at either end of a link.
[19:44] There's a transparent protocol that we've introduced that allows you to collect data from both sides of the link, which now makes the data you collect on either end more meaningful when you can put those two pieces together.
[19:54] Okay.
[19:56] Um let me get us back to the original part.
[20:00] Yes. You know, the 1.6 the 1.60 intercept is is critical, um but with you know, with the acquisition and eye on the future to what's going to happen with uh 3.2 terabit silicon photonics, whether that's a a CPO or 400 G per lane pluggables, um all of that's under our And and very very briefly, what timescales are we talking about 1.6 and that this year and 3.2 any I would say, you know, probably
[20:25] 3.2 any I would say, you know, probably we're looking in uh 2 years time.
[20:28] we're looking in uh 2 years time. >> Okay.
[20:31] So so we're seeing a a transition from
[20:33] So so we're seeing a a transition from from module-based solutions, right, to
[20:36] from module-based solutions, right, to to a near-package optics
[20:38] to a near-package optics um and then co-packaged optics
[20:40] um and then co-packaged optics to reduce power consumption, right,
[20:42] to reduce power consumption, right, going from from retime modules to linear
[20:45] going from from retime modules to linear NPO.
[20:46] NPO. Um and then CPO that will use chiplet
[20:48] Um and then CPO that will use chiplet technology and and die-to-die reducing
[20:51] technology and and die-to-die reducing the the power consumption even further.
[20:54] the the power consumption even further. And there's I mean, people who follow
[20:56] And there's I mean, people who follow those discussions, you know, at the big
[20:58] those discussions, you know, at the big conferences know that there's a lot of
[21:00] conferences know that there's a lot of debate
[21:01] debate when each of those crossover points are
[21:04] when each of those crossover points are occurring, right? So, I think the the
[21:06] occurring, right? So, I think the the crossover to NPO is is already
[21:08] crossover to NPO is is already happening, right? That's that's not a
[21:09] happening, right? That's that's not a secret. And then the when the the actual
[21:12] secret. And then the when the the actual crossover from NPO to CPU
[21:15] crossover from NPO to CPU will occur really in large volume for
[21:17] will occur really in large volume for the scale-up network, right? So it's so
[21:18] the scale-up network, right? So it's so not, you know, what what we've now seen
[21:21] not, you know, what what we've now seen uh more recently.
[21:23] uh more recently. Um so that's something we we're watching
[21:25] Um so that's something we we're watching very carefully. Okay.
[21:27] very carefully. Okay. And you're in stealth Sorry, I'm moving
[21:29] And you're in stealth Sorry, I'm moving along. You're in stealth mode and you've
[21:32] along. You're in stealth mode and you've you you're actually starting to
[21:33] you you're actually starting to demonstrate Aureole product and you've
[21:36] demonstrate Aureole product and you've just announced a second generation
[21:37] just announced a second generation product. So so both of you in terms of
[21:40] product. So so both of you in terms of Intercepts, what what you're aiming to
[21:42] Intercepts, what what you're aiming to When do you feel is a good time to hit
[21:44] When do you feel is a good time to hit the market and and with what? Yeah, I
[21:46] the market and and with what? Yeah, I think the LPO is something interesting
[21:48] think the LPO is something interesting also for the photonics part a lot
[21:50] also for the photonics part a lot because that probably is very important
[21:53] because that probably is very important when you have this end-to-end
[21:54] when you have this end-to-end communication having some uh switch in
[21:57] communication having some uh switch in the middle is photonics. It really
[21:59] the middle is photonics. It really general It's not trivial to deal with
[22:01] general It's not trivial to deal with it. So a lot of things develop by yeah,
[22:04] it. So a lot of things develop by yeah, also the DSP. That's very very
[22:06] also the DSP. That's very very interesting point and how to deal it.
[22:08] interesting point and how to deal it. Interesting to have this telemetry like
[22:10] Interesting to have this telemetry like you zero flop. It's good to hear about.
[22:12] you zero flop. It's good to hear about. And as well also the very interesting
[22:15] And as well also the very interesting later on to meet how much open will be
[22:17] later on to meet how much open will be the DSP to the users because that's
[22:20] the DSP to the users because that's maybe very very interesting here, right?
[22:23] maybe very very interesting here, right? And that I think is a very important
[22:24] And that I think is a very important because that that things is an important
[22:26] because that that things is an important a link where a disruption is made. Have
[22:29] a link where a disruption is made. Have a photonic switch make a disruption.
[22:31] a photonic switch make a disruption. That means it's not so easy to just say,
[22:33] That means it's not so easy to just say, "Okay, let's start and stop start and
[22:35] "Okay, let's start and stop start and stop whatever you want." So and that's
[22:37] stop whatever you want." So and that's what I mean. That's one of the challenge
[22:39] what I mean. That's one of the challenge to be solved. It's really to And and
[22:41] to be solved. It's really to And and it's not a trivial things to do. So
[22:43] it's not a trivial things to do. So that's what we are looking more uh as an
[22:45] that's what we are looking more uh as an interception to to move further in some
[22:47] interception to to move further in some of this technology.
[22:51] And as you said, yeah, we we actually
[22:52] And as you said, yeah, we we actually recently announced um the the the my the
[22:55] recently announced um the the the my the move up to scale-up architectures as
[22:58] move up to scale-up architectures as well, which comes largely from the
[22:59] well, which comes largely from the conversations we've had where people
[23:01] conversations we've had where people have said, "The energy reduction is
[23:03] have said, "The energy reduction is interesting, but actually we know we can
[23:05] interesting, but actually we know we can start to really drive up the bandwidth
[23:06] start to really drive up the bandwidth and throughput and so on." So, I think
[23:08] and throughput and so on." So, I think from from the And that's just in the
[23:10] from from the And that's just in the last year. Is that fair? Is 18 months?
[23:13] last year. Is that fair? Is 18 months? Yeah, there's there's a lot of changes
[23:16] Yeah, there's there's a lot of changes and this is such a fast-moving market,
[23:17] and this is such a fast-moving market, right? I think when we we first started
[23:20] right? I think when we we first started 3 years ago, it was really all about
[23:22] 3 years ago, it was really all about trading, that sort of thing. Now
[23:24] trading, that sort of thing. Now everyone's in more interested in
[23:25] everyone's in more interested in inference. Again, the the commonality
[23:27] inference. Again, the the commonality there is in in terms of uh
[23:30] there is in in terms of uh lower power, higher performance, and so
[23:32] lower power, higher performance, and so on. Um and and the the idea that if I'm
[23:34] on. Um and and the the idea that if I'm if I'm trying to push electrons around
[23:36] if I'm trying to push electrons around it, I've got a ceiling I'm going to hit.
[23:38] it, I've got a ceiling I'm going to hit. And [snorts] so, what can we do to
[23:39] And [snorts] so, what can we do to remove that ceiling and move beyond
[23:41] remove that ceiling and move beyond that?
[23:42] that? And I think the the idea about scaling
[23:44] And I think the the idea about scaling sideways, driving more and more
[23:45] sideways, driving more and more bandwidth, more and more
[23:47] bandwidth, more and more scale as well around that, lets us move
[23:49] scale as well around that, lets us move and think differently about where that
[23:51] and think differently about where that will go. So, I think whereas in the
[23:53] will go. So, I think whereas in the current product we're looking at
[23:55] current product we're looking at um yeah, we're we're demonstrating it to
[23:56] um yeah, we're we're demonstrating it to people right now. We're expecting
[23:57] people right now. We're expecting deployments for real later this year. Um
[24:00] deployments for real later this year. Um but really interesting to see where that
[24:02] but really interesting to see where that will be going in the next few years as
[24:03] will be going in the next few years as well.
[24:04] well. Okay, so that leads nicely onto the next
[24:06] Okay, so that leads nicely onto the next question, which was really one of the
[24:07] question, which was really one of the themes that Richard thought was really
[24:10] themes that Richard thought was really important with this panel.
[24:12] important with this panel. Which is just to understand the end
[24:14] Which is just to understand the end customers. So,
[24:15] customers. So, um both of you, the startups, are
[24:17] um both of you, the startups, are offering something quite a bit different
[24:19] offering something quite a bit different to what's being done today.
[24:21] to what's being done today. So, and you're startups. So, how do you
[24:24] So, and you're startups. So, how do you get this into the market? Who are you
[24:26] get this into the market? Who are you aiming it at? Is it specialist markets?
[24:29] aiming it at? Is it specialist markets? Is it hyperscalers? What type of
[24:31] Is it hyperscalers? What type of hyperscalers? How do you even get them
[24:33] hyperscalers? How do you even get them interested in something which is
[24:34] interested in something which is radically different to what they've been
[24:36] radically different to what they've been doing themselves?
[24:38] doing themselves? Um
[24:39] Um So, if if you think of the way that that
[24:41] So, if if you think of the way that that we're trying to do things with a passive
[24:43] we're trying to do things with a passive optics, it looks very similar to optical
[24:45] optics, it looks very similar to optical circuit switching, OCS. And what what
[24:48] circuit switching, OCS. And what what catches the attention when they realize
[24:49] catches the attention when they realize that uh we can actually deliver
[24:51] that uh we can actually deliver nanosecond scale switching, rapid
[24:54] nanosecond scale switching, rapid reconfiguration. We start to switch
[24:57] reconfiguration. We start to switch light [snorts] packets at the same speed
[24:59] light [snorts] packets at the same speed as you switch electrical packets, which
[25:01] as you switch electrical packets, which gets very interesting. So, we're seeing
[25:02] gets very interesting. So, we're seeing interest within hyperscalers that want
[25:04] interest within hyperscalers that want to do that. We're seeing interest within
[25:05] to do that. We're seeing interest within neo clouds that want to build specialist
[25:08] neo clouds that want to build specialist areas and then and workloads and so on.
[25:10] areas and then and workloads and so on. But actually, some of the verticals as
[25:12] But actually, some of the verticals as well, uh, financial technologies,
[25:14] well, uh, financial technologies, fintech and so on. Uh, high-frequency
[25:16] fintech and so on. Uh, high-frequency trading, they're looking for low
[25:18] trading, they're looking for low latency,
[25:19] latency, high speed, um, and and the ability to
[25:21] high speed, um, and and the ability to drive up in small focused workloads.
[25:24] drive up in small focused workloads. Now, whilst we can scale to very, very
[25:25] Now, whilst we can scale to very, very large networks, it's quite interesting.
[25:27] large networks, it's quite interesting. At the other end of the scale up, the
[25:29] At the other end of the scale up, the the size of the network is is more
[25:31] the size of the network is is more modest, thousands of GPUs, that sort of
[25:32] modest, thousands of GPUs, that sort of thing. But actually, it becomes much
[25:34] thing. But actually, it becomes much more focused on how fast you can deliver
[25:36] more focused on how fast you can deliver it. So, we're seeing it, um, from from
[25:39] it. So, we're seeing it, um, from from starting with a very focused idea of
[25:41] starting with a very focused idea of what we could make, we're able to show
[25:43] what we could make, we're able to show people how we'll re-implement it and
[25:44] people how we'll re-implement it and express it in different ways for a wider
[25:46] express it in different ways for a wider market. So, it's interesting times.
[25:50] market. So, it's interesting times. Yeah.
[25:51] Yeah. I cannot go to much specific here in the
[25:52] I cannot go to much specific here in the question, sorry. One thing is that one
[25:55] question, sorry. One thing is that one though just to mention, it is important
[25:57] though just to mention, it is important to be compliant with legacy.
[26:00] to be compliant with legacy. That's something that It is something
[26:02] That's something that It is something that's very important. Okay.
[26:05] that's very important. Okay. Well, from the perspective of a big
[26:07] Well, from the perspective of a big company, a little bit different, um,
[26:09] company, a little bit different, um, but so we focus very much on the on the
[26:11] but so we focus very much on the on the big hyperscalers,
[26:13] big hyperscalers, and I would say the way we approach this
[26:16] and I would say the way we approach this is to really make it a deep dialogue,
[26:18] is to really make it a deep dialogue, right? So, we we show what we can do,
[26:20] right? So, we we show what we can do, but we spend
[26:22] but we spend at least as much time, if not more,
[26:24] at least as much time, if not more, listening what they say they need and
[26:27] listening what they say they need and when they need it.
[26:28] when they need it. And to not over-engineer also and have
[26:31] And to not over-engineer also and have the [clears throat] time to market,
[26:32] the [clears throat] time to market, right? So, basically to be to have the
[26:34] right? So, basically to be to have the solution that they say they need at the
[26:36] solution that they say they need at the right moment.
[26:37] right moment. Um, and I think that's that's very
[26:39] Um, and I think that's that's very important to to stay I I
[26:41] important to to stay I I I noticed the open AI, uh, sorry, the
[26:44] I noticed the open AI, uh, sorry, the Anthropic, uh,
[26:46] Anthropic, uh, announcement with Amazon. That had a
[26:48] announcement with Amazon. That had a significant effect on your share price
[26:50] significant effect on your share price just recently.
[26:51] just recently. Just that announcement. Yeah.
[26:55] Just that announcement. Yeah. And and the same to to
[26:57] And and the same to to probably a similar answer. Definitely
[27:00] probably a similar answer. Definitely hyper scalar focus. I mean the the size
[27:02] hyper scalar focus. I mean the the size and scale that they're deploying at
[27:04] and scale that they're deploying at means they're usually the first have the
[27:06] means they're usually the first have the most demanding environments and the
[27:08] most demanding environments and the newest problems that we have to tackle.
[27:10] newest problems that we have to tackle. Um
[27:11] Um So that's a big part of our customer
[27:12] So that's a big part of our customer base. How much notice or head head start
[27:15] base. How much notice or head head start do you get How how much window do they
[27:17] do you get How how much window do they give you?
[27:19] give you? I mean in terms of you know preparation
[27:22] I mean in terms of you know preparation for the
[27:23] for the >> problems that are coming and and so that
[27:25] >> problems that are coming and and so that you can you and your engineers can start
[27:27] you can you and your engineers can start working on it or thinking of solutions.
[27:29] working on it or thinking of solutions. You know, I wish it was a lot more, but
[27:31] You know, I wish it was a lot more, but I think you know, what we're we're
[27:32] I think you know, what we're we're seeing now is problems occur and
[27:35] seeing now is problems occur and solutions are needed now. Nobody wants
[27:37] solutions are needed now. Nobody wants to wait till next year or 2 years to you
[27:40] to wait till next year or 2 years to you know, fix something. Um that's just the
[27:41] know, fix something. Um that's just the state of the industry.
[27:44] state of the industry. Uh one thing I will add about another
[27:46] Uh one thing I will add about another sort of group of customers that we're
[27:48] sort of group of customers that we're engaging with a lot the the new clouds.
[27:51] engaging with a lot the the new clouds. They generally don't have the same depth
[27:54] They generally don't have the same depth of you know
[27:55] of you know bench in their in you know, they don't
[27:57] bench in their in you know, they don't have a group of PhDs who are studying
[27:59] have a group of PhDs who are studying optics and like the hyper scalars do and
[28:01] optics and like the hyper scalars do and they they're often very knowledgeable.
[28:03] they they're often very knowledgeable. Neo clouds are trying to run really fast
[28:05] Neo clouds are trying to run really fast with a small team. They got lots of
[28:06] with a small team. They got lots of capital. So giving them telemetry and
[28:09] capital. So giving them telemetry and extra tools built into the the optics
[28:12] extra tools built into the the optics and pluggables has been very helpful to
[28:14] and pluggables has been very helpful to get them off the ground quickly. Okay,
[28:16] get them off the ground quickly. Okay, so I realize I'm I'm using most of the
[28:18] so I realize I'm I'm using most of the time. If I can throw this open, there's
[28:20] time. If I can throw this open, there's an opportunity to ask a really good
[28:21] an opportunity to ask a really good panel here.
[28:24] panel here. Sure.
[28:32] Hi guys. Um very interesting discussion.
[28:35] Hi guys. Um very interesting discussion. So when you look at the IO side of
[28:38] So when you look at the IO side of things, you you
[28:39] things, you you you moving data, lot of energy. So
[28:41] you moving data, lot of energy. So energy is the big problem.
[28:43] energy is the big problem. The the promise of optics photonics is
[28:46] The the promise of optics photonics is is a massive bandwidth scale up. But the
[28:49] is a massive bandwidth scale up. But the massive bandwidth scale up is blocked by
[28:51] massive bandwidth scale up is blocked by the sheer energy of that you take up
[28:53] the sheer energy of that you take up more and more more power than the
[28:55] more and more more power than the compute on the the thing you're trying
[28:57] compute on the the thing you're trying to get off of, right?
[28:58] to get off of, right? So, silicon photonics has been quite
[29:01] So, silicon photonics has been quite challenging. It's still at 5 pJ per bit,
[29:03] challenging. It's still at 5 pJ per bit, but if you go a long distance, it's it's
[29:05] but if you go a long distance, it's it's a big win.
[29:07] a big win. Uh the micro LED stuff is superb because
[29:10] Uh the micro LED stuff is superb because you can get sub pJ per bit
[29:12] you can get sub pJ per bit off of that package.
[29:14] off of that package. And so, I think it's a like you said, it
[29:16] And so, I think it's a like you said, it was a continuum.
[29:17] was a continuum. But it's
[29:18] But it's it
[29:20] it I I guess
[29:22] I I guess I guess that dichotomy of of the energy
[29:25] I guess that dichotomy of of the energy versus the bandwidth scaling is you
[29:26] versus the bandwidth scaling is you can't you can't get the bandwidth
[29:27] can't you can't get the bandwidth [clears throat] scaling unless you can
[29:29] [clears throat] scaling unless you can sort the power out first. And I think
[29:30] sort the power out first. And I think that's been silicon photonics' biggest
[29:32] that's been silicon photonics' biggest challenge, in all honesty.
[29:34] challenge, in all honesty. >> [clears throat]
[29:34] >> [clears throat] >> So, I was interesting your views on
[29:36] >> So, I was interesting your views on this.
[29:38] this. I I I would maybe frame the the problem
[29:40] I I I would maybe frame the the problem just a little bit differently because I
[29:42] just a little bit differently because I I think the way you put it, it sounds
[29:43] I think the way you put it, it sounds very much like replacing copper by fiber
[29:46] very much like replacing copper by fiber within a rack, right? And beating it by
[29:48] within a rack, right? And beating it by driving down power consumption. But I
[29:50] driving down power consumption. But I think what we're seeing is also that it
[29:51] think what we're seeing is also that it becomes increasingly difficult to scale
[29:54] becomes increasingly difficult to scale what a single rack can do because of
[29:56] what a single rack can do because of power consumption, right? Because you
[29:58] power consumption, right? Because you cannot cool more. So, the systems are
[30:00] cannot cool more. So, the systems are going to multi-rack.
[30:02] going to multi-rack. And at that point,
[30:03] And at that point, the the case for photonics becomes much
[30:07] the the case for photonics becomes much more compelling, right? Because it's not
[30:09] more compelling, right? Because it's not about incremental power consumption at
[30:11] about incremental power consumption at that point anymore. It's just that
[30:13] that point anymore. It's just that copper cannot do it, right? At the high
[30:14] copper cannot do it, right? At the high speeds if you you know, if you go across
[30:16] speeds if you you know, if you go across that in racks, for example, and and and
[30:18] that in racks, for example, and and and long distances. So, it's tied to power
[30:21] long distances. So, it's tied to power consumption, but it's
[30:23] consumption, but it's And of course, we have to reduce it as
[30:25] And of course, we have to reduce it as much as possible. But the this this just
[30:28] much as possible. But the this this just the the size of the system plays, I
[30:30] the the size of the system plays, I think, a major role in in the change we
[30:32] think, a major role in in the change we see.
[30:35] Yeah.
[30:36] Yeah. You know, maybe I'll add one comment on
[30:38] You know, maybe I'll add one comment on there. Um
[30:40] there. Um what we've seen with with liquid cooling
[30:41] what we've seen with with liquid cooling and a lot of folks you know moving to
[30:43] and a lot of folks you know moving to that approach to to actually cool these
[30:45] that approach to to actually cool these high density racks it's actually enabled
[30:47] high density racks it's actually enabled more use of copper and short reach
[30:49] more use of copper and short reach interconnects whether it's micro LED
[30:51] interconnects whether it's micro LED because you can jam more into a rack if
[30:53] because you can jam more into a rack if you have twice as many or 10 times as
[30:55] you have twice as many or 10 times as many GPUs connected in one rack because
[30:57] many GPUs connected in one rack because they're liquid cooled you know it allows
[30:59] they're liquid cooled you know it allows the use of some of the
[31:01] the use of some of the micro LED technology or or copper.
[31:05] One very quick last question or we just
[31:07] One very quick last question or we just very quick please just cuz we've been
[31:09] very quick please just cuz we've been signal that we're out of time but
[31:10] signal that we're out of time but please. Yes very quick one so
[31:13] please. Yes very quick one so we talk about a lot of a CPU in the
[31:15] we talk about a lot of a CPU in the scale out OCS potentially in the scale
[31:17] scale out OCS potentially in the scale out. To the to the microphone please so
[31:19] out. To the to the microphone please so everyone can hear. Yes so
[31:21] everyone can hear. Yes so scale up the CPU in the scale scale up
[31:25] scale up the CPU in the scale scale up OCS potentially for the scale out
[31:26] OCS potentially for the scale out coherent for the scale cross do you
[31:28] coherent for the scale cross do you really see a potential path that they're
[31:30] really see a potential path that they're having the full optical photonic based
[31:32] having the full optical photonic based architect for the future AI networking?
[31:35] architect for the future AI networking? Or is it just a you know
[31:38] Or is it just a you know segregated or segmented solutions for
[31:41] segregated or segmented solutions for different part of solutions or is there
[31:42] different part of solutions or is there any true
[31:44] any true end to end optical path? I I think in in
[31:47] end to end optical path? I I think in in the interim coexistence is really
[31:49] the interim coexistence is really important right exactly as as as my
[31:51] important right exactly as as as my friend here said you you need to be able
[31:53] friend here said you you need to be able to show that within a single rack I can
[31:55] to show that within a single rack I can drive copper then between racks I can do
[31:56] drive copper then between racks I can do drive photonics but we see that that
[31:59] drive photonics but we see that that progression as as
[32:01] progression as as there's going to be as many changes in
[32:02] there's going to be as many changes in the next couple of years as we've seen
[32:04] the next couple of years as we've seen so far. If you look at the a couple of
[32:06] so far. If you look at the a couple of years ago I think people were projecting
[32:08] years ago I think people were projecting about 25% of the energy was in the
[32:10] about 25% of the energy was in the networking
[32:11] networking projecting forward is it's 50 60% as you
[32:14] projecting forward is it's 50 60% as you get to terabits your tens of terabits
[32:17] get to terabits your tens of terabits between GPUs and and so on there has to
[32:19] between GPUs and and so on there has to be a change to that. Ultimately I think
[32:21] be a change to that. Ultimately I think as the previous presentation was is
[32:23] as the previous presentation was is going to start to see optics flowing
[32:25] going to start to see optics flowing right through the system and at that
[32:27] right through the system and at that point you need to be able to make sure
[32:29] point you need to be able to make sure you've got the full ecosystem that the
[32:31] you've got the full ecosystem that the transceivers that that switches and the
[32:33] transceivers that that switches and the rest of it but from our perspective we
[32:35] rest of it but from our perspective we believe you can eliminate
[32:36] believe you can eliminate electro-optical conversions, which is
[32:38] electro-optical conversions, which is where the power comes, and you can
[32:40] where the power comes, and you can change that.
[32:41] change that. If I can just add one more comment
[32:43] If I can just add one more comment before we wrap up is is that first of
[32:45] before we wrap up is is that first of all every hyperscaler at the moment has
[32:47] all every hyperscaler at the moment has its own system considerations. So people
[32:50] its own system considerations. So people do things differently. And the the
[32:52] do things differently. And the the challenge for everyone on this panel is
[32:54] challenge for everyone on this panel is to try and fit in and accommodate what
[32:56] to try and fit in and accommodate what they want. But where is this all going?
[32:59] they want. But where is this all going? And when I say where is it all going, it
[33:00] And when I say where is it all going, it takes time, but where is it going in 10,
[33:03] takes time, but where is it going in 10, 20, 30 years time? Well,
[33:04] 20, 30 years time? Well, the I make chart started to show where
[33:06] the I make chart started to show where you've got optics right next to the
[33:08] you've got optics right next to the chips and even wafer-based chips and
[33:10] chips and even wafer-based chips and waveguides.
[33:12] waveguides. That's some distance out. There's plenty
[33:14] That's some distance out. There's plenty of problems to solve in real real
[33:15] of problems to solve in real real commercial systems. But ultimately
[33:17] commercial systems. But ultimately that's optics. It just seems to be
[33:19] that's optics. It just seems to be encroaching everywhere.
[33:21] encroaching everywhere. So just So so
[33:24] So just So so it's valid what everything's said, but
[33:26] it's valid what everything's said, but what you've said. But ultimately that's
[33:28] what you've said. But ultimately that's the way it's going, but it may not be in
[33:30] the way it's going, but it may not be in in my lifetime.
[33:32] in my lifetime. Anyway, I want to thank you all and I
[33:34] Anyway, I want to thank you all and I want to thank Richard and I want to
[33:35] want to thank Richard and I want to thank the panel. It's been a really
[33:37] thank the panel. It's been a really great morning. So thank you all.
[33:40] great morning. So thank you all. >> [applause]