# @HPCpodcast-54: Silicon Photonics, w Keren Bergman

https://www.youtube.com/watch?v=9wxX_GIGmMw

[00:04] Photonix doesn't need the latest node.
[00:06] In fact, it's not even advantageous.
[00:10] When we talk about photonics, we really mean integrated optics.
[00:12] Optics that's put on a chip and co-integrated with electronics.
[00:20] The broad idea that in the optical domain, we can send multiple signals together in the same channel.
[00:25] extremely high bandwidth densities like bandwidth densities of on the orders of multiple terabits per millimeter.
[00:32] In fact, in my own work, we just have a record chip that we made that has over 5 terabs per millime squared.
[00:38] That's on the research side.
[00:41] From Orion X in association with inside HPC.
[00:44] This is the ATH HPC podcast.
[00:47] Join Shaheen Khan and Doug Black as they discuss supercomputing technologies and the applications, markets, and policies that shape them.
[00:52] Thank you for being with us.
[00:54] Hi Shaheen.
[00:56] It's great to be with you again as always.
[00:58] So excited to be here today once more because we have a wonderful special guest.
[01:00] We do.
[01:02] We have with us Karen Bergman.
[01:02] She is the
[01:05] Charles Bachelor professor of electrical engineering at Columbia University and her area of research is really at the intersection of photonics and computing.
[01:16] And the topic for today is photonix. its potential possibly to transform system performance with a much faster and also much more energyefficient interconnect for data movement.
[01:30] So Karen, welcome.
[01:30] Thank you so much. I'm excited to be part of this podcast. Looking forward to talking about my favorite topics.
[01:37] Great. I'm really excited too because the last two years coming out of the supercomputing conference, I've written articles about panels that you've been on on photonics and to hear the photonix community talk about the potential of this technology.
[01:49] Really tremendous tremendous potential.
[01:53] Karen, maybe we could start off talk about the problem that photonix is trying to address visav existing copperbased interconnects and some of the consequences of that technology that possibly is really being
[02:06] pushed to its limits within the context
[02:09] of computing systems which is really
[02:11] like the kind of focus on that perhaps
[02:14] you know one of the most important
[02:15] challenges we have is is data movement.
[02:17] So data movement costs a lot of energy
[02:20] and it's also difficult in the
[02:23] electronic domain to support large
[02:26] volumes of data high bandwidth densities
[02:29] especially when we try to propagate that
[02:31] data over longer distances. So in
[02:34] current electronics current processor
[02:37] chips memory chips we have excellent
[02:40] electronic interconnects. We've had them
[02:42] for you know for decades. We can
[02:44] integrate them with very high densities
[02:47] and as long as the distance over which
[02:50] we want to move that data basically
[02:53] within very much within the chip is
[02:55] short the energy consumption is very
[02:58] modest in the phentojoules per bit
[03:00] numbers. But the challenge is that as we
[03:03] try to move data electronically over
[03:06] longer distances, the energy that's
[03:08] required to move that data, especially high bandwidth data, high frequency data, uh just increases and increases nonlinearly and it's primarily because of of the losses that we have in propagating signals, high frequency signals over distances and they need to be reamplified and and and so forth.
[03:27] So what exists today in computing systems broadly speaking is that within the chip or even within the the socket or the interposer we can have pretty good high density communications high bandwidth with also modest energy consumption good high energy efficiency but as soon as we try to move that data any appreciable distance out of the socket off the chip out of the package the energy consumption just increases tremendously and and So therefore in today's systems the amount of bandwidth that's available out of the package is very limited.
[04:03] It's typically two orders of magnitude less than what we have inside the socket or
[04:08] the package.
[04:10] And so going back to the computer systems and and what we're trying to do we're trying to scale applications.
[04:13] We're trying to run high performance especially you know today's systems are running a lot of AI uh distributed machine learning applications.
[04:23] it's increasing by an order of magnitude per year in terms of the model sizes.
[04:28] So as we try to create and there's a limit to how much of the problem you can put within one package right uh so we have to scale the system to large data centers large high performance computing systems and so as we interconnect or the communication among these computing uh nodes typically it would have a computing an accelerator a memory within that socket the communication between them is is really really limited it's kind of like these computing nodes are trying to drink through a through a tight straw and that is limiting both the efficiency of running the application.
[05:01] It's limiting how much we can continue to scale the application and of course the overall power consumption.
[05:07] So where optics is
[05:10] coming in or more correctly photonix.
[05:13] the difference is that when we talk about photonics we really mean integrated optics optics that's put on a chip and co-integrated with electronics.
[05:20] So there is a a big desire to bring the fatonic interface to that socket to where that compute node memory node exists and be able to then use the optical domain to propagate the data over the longer distances.
[05:40] And so that's really the the main challenge that or the promise of this this new technology that's emerging is to really open up the communication among the compute nodes or the the nodes that are used to communicate in in these large scale systems.
[05:58] Excellent.
[06:00] Karen, you mentioned the chip level on the socket all the way down to the miniature level.
[06:05] The average person's interactions with optical is where their you know network provider pulls fiber optics to
[06:12] their house and now they have faster communication across the country across the globe.
[06:17] What are the differences between working at the telecommunications level compared to when you're working down at the chip level and potentially everything in between?
[06:26] In what way does the science and the technology challenges change qualitatively?
[06:31] So, we're all familiar with fiber optics, you know, for decades.
[06:35] Some of us remember the old Sprint commercial with a pin drop.
[06:37] And we know that from our experiences that fiber optics have have revolutionized communication, telecommunication more specifically across the globe and allow us to, you know, have broad communications over over very long distances.
[06:53] You know, this has driven the telecom market for for decades.
[06:55] Now, we talk about phatonics that we're hoping to use within computing systems.
[07:01] And one big difference, of course, is that the distances are are shorter.
[07:04] Typically, we're we're talking about using optics to communicate not over long distances,
[07:14] transatlantic or trans-pacific, but within a data center or maybe even just between chips.
[07:23] So, those are uh essentially different technologies.
[07:24] At the core, of course, we're still talking about optics.
[07:29] So that means that we are going to be sending signals in the optical domain and many of the technologies that have been developed for the telecom market, fiber optic market certainly can be translated to these new emerging applications within computing systems.
[07:45] But some of the technologies are not necessarily as translatable.
[07:48] So let me go one step a little deeper.
[07:50] In the fiber optic domain for example we use something that we call wavelength division multiplexing.
[07:57] That means that within one fiber we can send multiple signals.
[08:03] Each signal is put on a different optical wavelength.
[08:05] It's called WDM for the division multiplexing and that enables us to have very high bandwidth that we can send over a single fiber.
[08:13] So the same idea
[08:16] also applies to how we would communicate optically with photonics on on a chip.
[08:21] And I'll I'll explain you know how we might be able to do that.
[08:23] But just the the broad idea that in the optical domain we can send multiple signals together in the same channel in the same either it's a fiber optic waveguide or it's a a waveguide that's sitting on on a silicon chip but it's one link one medium where we can send multiple signals at the same time.
[08:42] And that is really the one of the key enablers of the high bandwidth density that we can achieve in optical communications.
[08:48] One follow-up question.
[08:51] You mentioned wavelength division multiplexing.
[08:54] When you look at a light wave, there's also amplitude and frequency and phase and all other aspects of light that you know if you're not close to it, you wouldn't even know they exist.
[09:04] Are all also involved and eligible to be controlled or is wavelength really the way to go and has proven to be?
[09:09] How does that world work?
[09:14] Sure. So, first we when we
[09:16] started fiber optics decades ago, it was just a single channel and then of course wavelength division multiplexing became the next kind of revolution in in fiber optic communication and really enabled by you know these what are known as the herbium doped amplifiers.
[09:30] And more recently as the bandwidth demands continue to grow and grow the other components of the light phase in particular and and so forth are being used to increase further increase the bandwidth that are carried by by that fiber.
[09:45] So in today's system again the telecom market coherent communication is being used widely and there's been you know tremendous advances in that side.
[09:54] On the other hand on the computing application that we're going to talk about more here we're just starting.
[09:58] So the first step is really going to be with wavelength division multiplexing.
[10:03] You know looking at coherent communication and so forth may come at a future date but right now you know the challenges are are really on trying to get the WDM links for this application.
[10:13] Now when you say coherent communication
[10:18] do you mean it in the sense of cache coherency where multiple players need to agree on something or is it more a question of quality?
[10:25] How's that defined in this context?
[10:27] Yeah, in the context of optical communications, coherent means that you're using the phase of the light and higher order modulation formats and so forth.
[10:37] So that's sort of a whole different dimension of scaling the bandwidths.
[10:38] Got it.
[10:40] Excellent.
[10:43] Karen, let me ask you maybe from a if not 30,000 foot level, 15,000 foot level, you know, where do we stand now in this overall push to bring photonics to commercial readiness and specifically I'm really interested in, you know, what there's a lot of skepticism in our community, the HPC community about this technology.
[11:05] I did note in the article I wrote coming out of the supercomputing conference that the photonix panel that you were a part of on the very last day toward the end of the very last day of the conference.
[11:15] Um yet I will say this
[11:18] The article I wrote got a lot of interest.
[11:21] So I think there's a lot of interest but maybe combined with skepticism.
[11:25] Where does that skepticism come from?
[11:27] What are the biggest challenges on the path to commercial readiness?
[11:31] So the our best hope right now in terms of let me start by saying that there's a strong interest in addressing the high level problem the data movement problem right so platonics is capable of just inherently physically we can put many wavelengths in the same in the same pin if you will in the same link and we can modulate each one of those wavelengths at at high speeds.
[11:53] So potentially we have the capability of just from the physics of being able to create an interconnect with extremely high bandwidth densities like bandwidth densities of on the orders of multiple terabits per millimeter.
[12:08] In fact in my own work we just have a record chip that we made that has over 5 terabs per millimeter squared.
[12:14] That's on the research side.
[12:15] In addition to that, because of the relatively low loss of the optical propagation medium, whether
[12:20] it's fibers or or interconnects on a chip,
[12:22] we can move that data over long distances without having to reamplify it.
[12:27] And so there is this kind of distance independence quality characteristic of optical communications,
[12:33] especially for that's important for for computing applications.
[12:37] So, so there's a big promise of being able to provide ubiquitous connectivity, high high bandwidth across the computer system with minimal energy consumption.
[12:46] So that's where we want to go.
[12:48] Now looking at where we are in the last about 15 years, there's been another revolution in the optical world and the revolution is in the field of silicon photonics.
[12:58] So, Vatonics in the past has been, you know, these large kind of macro scale lenses and things that you imagine on an optical table.
[13:09] And in the last, you know, since about the early 2000s, we've really pushed hard on bringing fatonics not just to the chip, but to silicon to
[13:21] CMOS fabrication.
[13:23] And this has been a globalwide effort with efforts within the US and and globally Asia and in Europe as well of using CMOS fabrication facilities to actually essentially fabricate integrated structures, photonic structures, modulators, receivers, transmitters, filters, you know, everything that you would in the past have on an optical bench all on a chip.
[13:49] And in the last about five years or so, I would say we've really come to maturation of that.
[13:52] So we now have the state-of-the-art is that we have 300 millimeter fabs including uh commercial ones like global foundaries, TSMC and others across the globe where you can basically submit photonic circuits, phatonic designs and they will get fabricated in the fab.
[14:09] So that's great.
[14:12] That was a big part of the commercialization ecosystem that was brought up.
[14:16] But now the main challenges on the technology side there are many challenges.
[14:18] One of the main ones is
[14:22] really on the assembly and the packaging.
[14:25] It's almost a little bit the opposite of electronics.
[14:27] Electronics the fab is expensive and then the packaging is cheap.
[14:31] In the optical world, platonic world, the fabrication is not that expensive compared to the packaging because you know there is it's still early days in terms of how do you package the fatonic chip with the electronic chip and how do you package the whole thing with the laser whether the laser could be on the chip off the chip there all kinds of different solutions out there and that's really where I think the ecosystem still needs to be developed.
[14:59] You know, there's a lot of efforts that are being pushed forward right now.
[15:03] And I will say optimistically that this is not just a research interest or advanced research that is out there.
[15:11] This is actually an area where really all the major systems vendors from Nvidia to AMD to Intel to HPE and and so on are investing.
[15:25] Seriously investing in silicon photonics.
[15:28] Because it's clear that this is the needed solution for interconnects in the future.
[15:32] We see the road map.
[15:34] We see the path that it's going to happen but today the challenges are what is the right packaging solution?
[15:39] What is the assembly?
[15:41] How is photonics going to be essentially integrated together within the compute system?
[15:46] Karen, is this fundamentally or substantially because we are mixing analog and digital and analog systems always need tuning and all that or is there more complexity to it in addition?
[15:57] Absolutely both.
[15:59] Yes.
[16:01] It's it's essentially optics is an analog signal.
[16:03] Yes, absolutely.
[16:05] And of course, you know, we modulate it, we digitize it, but it's essentially an analog circuit and it has other peculiar issues.
[16:11] In general, optics is or photonics, I should I'm using the words interchangeably, but I mean photonics, anything optical is going to be thermally sensitive because as soon as you raise the temperature or change the temperature, the index of refraction of the material.
[16:27] changes and whatever you have in that material will be sensitive to that.
[16:31] So solutions for on the thermal side and the packaging side and then the other added complexity is of course the laser but the other added complexity is that there isn't necessarily a standard interface that we are all designing towards right now in terms of the electronic fatonic interface right that's a really key issue there's been a lot of discussions you know is it going to be something like CXL or other kinds of protocols but that's another we still don't have kind of a standard packaging solution, assembly solution, interface solution that everyone is designing towards.
[17:05] It's a bit of a open field at the moment.
[17:08] There's a lot of effort to try to get there, but that's I would say one of the key challenges.
[17:14] Maybe this is a segue into photonics for communication versus computation and the mixed signal chips that are being designed traditionally for communication substantially, but now maybe increasingly for computation.
[17:26] And there
[17:29] are startups that are pointing to optical computation in a classical physics way sort of a way.
[17:35] How does that change the complexity of all of this?
[17:37] And you also mentioned interposer early in our conversation and that also seems to be something that only shows up if you're talking about chiplets and you know way down at one end of the spectrum.
[17:48] Are they morphing?
[17:50] Should we expect to see more and more optical computing in addition to communication?
[17:54] I don't think so.
[17:56] I mean this might be a bit a bit of a controversial viewpoint but I really don't think that there is a lot of in my opinion promise for optical computing per se like really doing computation in the optical domain.
[18:07] Optics is is really not a great technology for computation for various reasons.
[18:12] You know, transistors are are really really good for computation and they have kind of the nonlinear function and all of that and doing it in the optical domain is has been kind of swimming against the tide of of physical nature of optics or or photonics.
[18:27] There are special cases like you can imagine.
[18:30] So for example, there's the old idea of doing an optical FFT in the spatial domain.
[18:34] And you can carry that over now with new technology and doing things like that on a chip and doing shuffling and different kinds of data movements.
[18:42] And it's interesting, it's curious and you can maybe even show that you can accelerate that part of the computation.
[18:49] But when you put it all together, at the end of the day, you need to have a system around that for that little accelerator or that little component to be useful.
[18:58] When you put the whole system around it and work out what the numbers are in terms of really have you accelerated the entire application?
[19:07] In the work that I've done the answer is no.
[19:09] So I know that there are companies and there are other researchers working in this direction but in my opinion it it's not necessarily the right.
[19:15] However for interconnects for communications and for networking which would include switching I believe that it is the right technology and we it's kind of been proven in the past in just look at fiber optics and how success uccessful that's been.
[19:27] So for data movement and
[19:31] communication, optics is really the very much very promising solution.
[19:36] So Karen, earlier today I sent over to you an excerpt from a paper from Satoshi Matsuaka at the Reichen Institute and colleagues on the myths and legends of HPC.
[19:46] paper came out earlier this year and in the context of disagregated computing he brought up reservations and questions in regard to silicon photonics.
[19:58] and he specifically he focused on lowcost manufacturing which you have talked about but also the switching issue optical switching versus circuit switching.
[20:07] and you seem to be saying that to go with photonix we servers or systems would have to move toward circuit switching which is contrary to where things are today.
[20:19] Could you address that issue?
[20:22] Of course, this is a a really nice paper.
[20:23] I'm glad you shared it with me and I basically agree with what he's saying, but with a caveat.
[20:29] So, let let me explain.
[20:31] Absolutely. Phetonics or optics optical
[20:33] interconnection networks specifically
[20:34] for computing the switching is going to
[20:36] be circuit switching because trying to
[20:39] implement packet switching in the
[20:41] optical domain is essentially a fool's
[20:43] errand. I've tried to do it. I've done a
[20:45] lot of research on it. I have papers on
[20:46] it. But the bottom line is that you have
[20:49] to really do let's say strictly packet
[20:52] level switching. You need to have
[20:54] buffers. You need to have some kind of
[20:56] storage capability even if small. And to
[20:59] do that in the optical domain is really
[21:01] challenging and it's fundamental.
[21:03] Storing the light and being able to take
[21:06] it out as you would like a electronic
[21:08] buffer, electronic ram. It's just like I
[21:11] said, it's kind of like swimming uphill.
[21:13] it really does not fit the the physical
[21:15] nature of the technology. So circuit
[21:18] switching however can be done and
[21:19] there's various technologies in fact
[21:21] even Google had a a paper and other
[21:24] people of course including myself in the
[21:26] past have have written done a lot of
[21:28] work on circuit switching within the
[21:30] context of data centers and computing
[21:33] and this is really where the big caveat
[21:34] is with the really important point it's
[21:36] not instead of the electronic packet
[21:39] switches it's in addition I think
[21:42] switching will be part of the solution I
[21:44] think initially it'll just be the
[21:45] interconnect because you can put the
[21:47] optical IO's just as the interfaces
[21:50] through the electronic switches. You
[21:51] have a let's say 100 terab electronic
[21:54] switch or even higher and you can put
[21:56] this highdensity optical IO right there
[21:58] and be able to have incredible high
[22:01] bandwidth connectivity in the system and
[22:03] all the benefits of electronic packet
[22:05] switches. But in addition to that, you
[22:08] might want to have these transparent
[22:10] connectivity lanes that enable you to
[22:13] connect from longer distances across the
[22:15] data center or even within the rack all
[22:17] optically. And so there are various
[22:19] solutions for adding circuit switches
[22:22] together with the electronic packet
[22:25] switches in these hybrid architectures.
[22:27] Very very interesting. One question.
[22:29] Photons are so fleeting they could
[22:32] disappear on you. So the effort taken to
[22:36] make sure that doesn't happen must cost
[22:38] some efficiency. Are we so much more
[22:40] efficient that that just in the noise
[22:42] and doesn't matter and to what extent
[22:43] are we reaching the theoretical limits
[22:46] of efficiency for data transfer? I guess
[22:48] I'm not completely clear on the
[22:50] question. Do you mean in terms of energy
[22:52] efficiency or the losses in the system?
[22:54] Really both. I think initially is really
[22:56] the loss in the system and what it takes
[22:58] to mitigate that. But also when you look
[23:01] at the information theory on what is the
[23:04] minimum amount of energy required to
[23:06] transfer some information from point A
[23:09] to point B and some thermodynamics
[23:11] analysis of that leads to some kind of a
[23:14] theoretical limit. Are we there yet? Are
[23:16] we close to it? Once we're done with
[23:18] optical communication, is this it? Is
[23:21] that as fast as it's going to get? Or
[23:23] are there other technologies behind it
[23:25] that might reemerge? That's the second
[23:27] part of that question. Understood. Thank
[23:28] you. So, we're not close to the limits.
[23:30] There's definitely way more loss. The
[23:33] silicon phetonics is great. The fiber
[23:35] optics are are great in terms of their
[23:37] their loss profiles, but there's a lot
[23:39] more to do. And again, it kind of goes
[23:41] back to the challenges of why isn't this
[23:44] already in the systems? Why aren't
[23:46] fatonic triplet IO's already part of
[23:48] commercial available products? And there
[23:51] are losses at the interfaces. There are
[23:53] losses still ways that need to be done
[23:56] to reduce the losses that we have today.
[23:58] So there's ways to go there. Definitely
[24:01] more that can be pushed on the energy
[24:03] efficiency for sure. I think in terms of
[24:06] fundamental limits, there's probably at
[24:07] least another one order if not two
[24:09] orders of magnitude that we can one
[24:11] order. I would say two orders would be
[24:14] really pushing it. But yeah, so there
[24:15] there's a lot more to do. Also, you
[24:17] know, the laser is also a big part of
[24:19] the story and there's tremendous amount
[24:21] of room to improve there in terms of its
[24:24] efficiency. Lasers are not efficient as
[24:27] a whole right now. Kind of a typical
[24:29] laser is sort of in the 5% if you're
[24:33] lucky wall plug efficiency to the light
[24:36] version. So it's really poor. Even with
[24:38] that photonic interconnects using these
[24:41] kind of poor laser efficient sources are
[24:44] still way more energy efficient than the
[24:47] electronic interconnect. So there's a
[24:49] lot of room to grow there for sure.
[24:50] Excellent. So one other question is
[24:52] really the supply chain of this emerging
[24:55] technology. You mentioned lasers that's
[24:58] kind of the optics. You look at who in
[25:00] the world scene is really good at this
[25:02] sort of a thing. How does the supply
[25:04] chain look for this kind of emerging
[25:06] technology? That's one of the major
[25:08] another one of the major challenges
[25:10] right the supply chain. So I mentioned
[25:12] on the actual photonic devices they
[25:14] integrated the ability to now have
[25:16] fabrication in 300 mm wafer scale
[25:20] process. So that's great and I think
[25:22] that that probably can be scaled up
[25:24] pretty readily, right? You can start
[25:25] producing on the laser side also. You
[25:28] know, there are definitely suppliers out
[25:30] there that have been supplying lasers
[25:32] for the telecom market and but there's a
[25:34] lot more that needs to be done there in
[25:37] terms of bringing the lasers either to
[25:39] the chip or solving the external to the
[25:41] chip packaging problem. And it's a bit
[25:43] of a catch22 because yes, all these
[25:46] things can be scaled and it's not kind
[25:49] of a fundamental technology challenge,
[25:51] but in order for them to scale, there
[25:54] has to be a market and in order for
[25:56] there to be a market, it has to be low
[25:57] enough cost. So we're kind of in this
[25:59] sometimes they call it the valley of
[26:01] death or something like that, right? In
[26:04] transition. So this is where we are
[26:06] right now, I believe. But luckily
[26:08] there's a lot of investment both from
[26:11] the commercial side themselves like I
[26:13] mentioned really all the major vendors
[26:15] have investments in this area but also
[26:17] from government entities and and the
[26:19] public sector. So obviously the chips
[26:22] act that's coming in the US will be I
[26:24] think one of the factors that helps
[26:26] propels this forward as part of the
[26:28] heterogeneous integration that will be
[26:30] heavily pushed in those efforts. I see.
[26:33] Now you mentioned the fabs and those are
[26:36] not the leading edge you know 2 three
[26:38] nanometer fabs they're in principle they
[26:41] should be easier to do and more common
[26:44] what do they have to do to be able to
[26:45] also incorporate silicon photonics you
[26:48] couldn't just go to any fab and have it
[26:51] done right they have to be outfitted to
[26:53] do that is that a lot of work and is it
[26:55] already done or it's done to some extent
[26:58] already it's not necessarily that they
[27:00] have to be outfitted but the process has
[27:02] to be developed you need to go through
[27:04] some process development, PDK
[27:06] development that is optimized for the
[27:08] phatonic devices to work that's going on
[27:11] to some extent like I mentioned global
[27:14] already has a commercial photonix
[27:17] process I know he is working on one so
[27:20] forth and then there are smaller fabs
[27:21] and as you said fatonix doesn't need the
[27:24] latest node in fact it's not even
[27:26] advantageous and so there's a lower cost
[27:29] with that with fabricating fatonic
[27:32] wafers versus like a 2 nanometer 3
[27:34] nanometer CMOS. So I I think I think
[27:36] it's on the fab side the capability is
[27:39] is pretty much there and it's just kind
[27:41] of waiting for when the the key is
[27:44] turned and you know volumes are being
[27:46] scaled. I think that part will be ready
[27:48] to go but you know the other parts is is
[27:50] where really the challenges remain and
[27:52] as you mentioned it makes up for it with
[27:54] packaging and assembly and is that all
[27:56] in the US where does that take place the
[27:58] assembly and packaging? Oh that's all
[28:00] over the world. So all the way over the
[28:02] world. Yeah. Yeah. Absolutely. I mean
[28:03] obviously a lot of the packaging
[28:05] assembly is done there's a lot in Asia,
[28:07] some in the US and Europe of course. So
[28:10] pretty much every globally there are
[28:12] efforts in this direction for sure. So
[28:15] Karen I guess my last question would be
[28:17] gazing into your crystal ball. the
[28:19] annoying question and possibly based on
[28:21] the rate of progress you've seen over
[28:24] the last 10 15 years when will we reach
[28:27] kind of commercial life where a lot of
[28:29] these challenges are overcome and we
[28:32] really see the adoption of of this
[28:33] technology. Yeah, great question of
[28:36] course. So the way I would answer it is
[28:38] adoption will happen when the cost goes
[28:40] down. Photonix is definitely giving us
[28:43] this better performance, better energy
[28:44] consumption. technology itself is proven
[28:46] been lots of prototypes and examples and
[28:49] that sort of work already out there. Now
[28:52] it's about scaling it up and reducing
[28:54] the cost. A lot of it as I mentioned is
[28:56] on the packaging side and at the same
[28:58] time you have the pain point of the
[29:00] electronic side and fortunately the pain
[29:03] point of the electronics is accelerating
[29:06] because thankfully we have all these
[29:08] distributed AI systems that are out
[29:10] there and and keep growing at this
[29:12] mindboggling pace and so the need for
[29:16] larger better data movement is just
[29:19] exploding. I think the turning point is
[29:21] sort of in the three to five year.
[29:23] That's my sense. I think in the near
[29:24] term it's going to be on the high end
[29:26] obviously like usually the adoption on
[29:29] the high-end systems and then uh soon
[29:31] after that once the manufacturing is
[29:33] turned on it'll be more widely adopted.
[29:35] So that that's my time frame. I think in
[29:37] the three to five year time frame. Okay.
[29:40] Brilliant. We just had a big conference
[29:42] in the photonics world just earlier this
[29:44] month. So this is a conference that's
[29:46] been traditionally telecom fiber optics
[29:49] and has gotten more and more into the
[29:51] photonic interconnects for computing as
[29:53] we've been talking about today and at
[29:56] this conference now we have speakers
[29:58] from Nvidia from Intel from right so all
[30:02] the I believe the turning point is
[30:05] closer than than it's ever been the
[30:07] unusual suspect that's
[30:09] right it's starting to look like SC you
[30:12] know right right well you Oh, isn't that
[30:15] a good testimony for SC that everything
[30:17] is going to end up being like it? Yeah,
[30:20] that's great. So, Karen, it's been great
[30:22] talking with you. Thank you for this
[30:24] overview of this really interesting
[30:26] technology and all the best of luck with
[30:29] your continuing research and thanks so
[30:31] much for being with us. Thank you so
[30:33] much and so fun. Like you said, I can
[30:35] talk about this forever.
[30:37] Excellent.
[30:39] That's it for this episode of the ATHC
[30:42] podcast. Every episode is featured on
[30:45] insideh.com and posted on orionex.net.
[30:48] Use the comment section or tweet us with
[30:50] any questions or to propose topics of
[30:52] discussion. If you like the show, rate
[30:54] and review it on Apple Podcasts or
[30:56] wherever you listen. The HPC podcast is
[30:59] a production of OrionX in association
[31:02] with Inside HPC. Thank you for
[31:04] listening.