Full Transcript
https://www.youtube.com/watch?v=WovVoqAe9Sk
[00:06] So welcome to the distinguishers lecture.
[00:08] So welcome to the distinguishers lecture series.
[00:10] Um just to remind you and also people online.
[00:13] So this is our lecture series um where we invite the main thought leaders um of their research fields um to come to our institute and tell us about their visions.
[00:23] And so it's my pleasure um to uh present you Karen Burkman.
[00:29] She's a a professor of electrical engineering at Colombia University and also a faculty director of Colombia nano initiative.
[00:37] Um she's also director of center for ubi which was connectivity by DARPA and she has won uh many prizes um amongst them recently the ITLE e um photonics engineering award and the optica ck um me medal um which are very high profile prices and I'm very happy to have you here at the institute actually Karen was here already 30 years ago I'm working as a summer intern and get loss
[01:06] a summer intern and get loss chair at that time at the university we.
[01:08] chair at that time at the university we found out yesterday evening.
[01:10] found out yesterday evening. So there's a connection to Alangan.
[01:13] Um so we are welcoming you and we are happy to hear.
[01:15] from about photonic connectivity for AI systems.
[01:39] Thank you very much. Um and uh so.
[01:44] beautiful place and of course to get to.
[01:46] see the new now for me new but not not so new uh Mox Blanc Institute it's it's wonderful really amazing.
[01:52] Okay uh this will be a little bit more systemsoriented talk.
[01:57] So um please feel free to interrupt if if there's any questions along the way.
[02:04] I don't mind to keep it keep it informal.
[02:08] Keep it keep it informal.
[02:12] So um I'm sure everyone is uh aware of the revolution that we're undergoing with AI systems.
[02:16] Uh many many charts like this uh uh have have been out there.
[02:19] I just want to point out a few things.
[02:21] Um so of course this is showing the amount of uh compute that's going into AI and I'm I'm only showing the training in this case because that's really the most compute intensive part of of the AI, right?
[02:35] Training training the big uh uh large uh learning models right the ALMs and you can see some familiar things over here right everybody knows GPT here is Gro DeepS right the Chinese uh uh algorithm that uh had a big impact a big splash uh bit about a year and a half ago.
[03:00] Uh so what's happening here is uh first let's let's look at the years so this is a very short period of time, right?
[03:08] a very short period of time, right?
[03:08] Uh about a decade or so.
[03:11] Um really there about a decade or so.
[03:11] Um really there was almost an inflection point between 2018 and 2020.
[03:14] was almost an inflection point between 2018 and 2020.
[03:18] So maybe eight years ago or so really things started to accelerate and I want to point out that the uh vertical axis is on a logarithmic scale.
[03:20] accelerate and I want to point out that the uh vertical axis is on a logarithmic scale.
[03:23] the uh vertical axis is on a logarithmic scale.
[03:27] This is a very short number of years logarithmic scale.
[03:29] scale. This is a very short number of years
[03:31] years logarithmic scale.
[03:34] Okay, this is this is a big big big deal.
[03:34] Um, so we've not seen anything at least in my career that has accelerated this much this quickly.
[03:38] seen anything at least in my career that has accelerated this much this quickly.
[03:42] has accelerated this much this quickly.
[03:45] So when everybody's talking about AI, you know, there's something to it.
[03:47] So when everybody's talking about AI, you know, there's something to it.
[03:49] I mean, there really we've not seen uh what what's happening is essentially things are increasing by about an order of magnitude every year.
[03:52] mean, there really we've not seen uh what what's happening is essentially
[03:54] what what's happening is essentially things are increasing by about an order of magnitude every year.
[03:56] things are increasing by about an order of magnitude every year.
[03:59] Um I I I don't know of many things in physics, science, technology that that have actually made gone this way.
[04:02] Um I I I don't know of many things in physics, science, technology that that
[04:05] physics, science, technology that that have actually made gone this way.
[04:05] Is it
[04:09] Have actually made gone this way.
[04:09] Is it slowing down?
[04:09] Not quite yet.
[04:09] Not yet.
[04:12] Slowing down?
[04:12] Not quite yet.
[04:12] Not yet.
[04:12] Not yet.
[04:14] Not yet.
[04:14] I mean definitely there's lots of changes.
[04:16] We're going more from training to the domain specific inference applications and you know many many things are going on but this is this is big.
[04:23] Um all right.
[04:23] Uh so let let's keep all of this in mind.
[04:29] And uh um what I'm interested in uh I'm a I'm a optical communications person fatonic interconnects and so we are interested in what's happening to the communication with these large AI systems you know and what can fatonix do uh to to uh impact impact these systems.
[04:49] All right.
[04:49] So this is going up very quickly.
[04:52] At the same time, another very important challenge um is the growth in the energy consumption associated with training these big LLMs, right?
[05:01] How much energy is being spent uh on on AI essentially?
[05:07] And of course, you've probably heard in
[05:09] And of course, you've probably heard in the news, you know, the the data centers.
[05:11] the news, you know, the the data centers that are being installed all over the world.
[05:13] You know, they're equivalent to the um you know, the all the energy of of like Ireland is given as as an example usually, you know, because it's a relatively small country.
[05:22] Um but the these are big big numbers in in power consumption and energy consumption.
[05:32] And you know, we think about computing and we're like, "Oh, we see this curve going up, orders of magnitude, then yay, computing is going, but energy is is not uh, you know, is is is not a sustainable uh resource, right?
[05:45] And uh I know we care a lot in aan about sustainability.
[05:51] I had a beautiful hotel where everything is sustainable.
[05:53] So, I felt very good about myself uh being here.
[05:55] Um, oh, actually all the way back to 30 years ago, the the the apartment where we stayed was uh solar powered.
[06:07] So, it was great.
[06:09] We had the little issue with the hot water.
[06:09] So, you you turn it on and uh
[06:12] hot water.
[06:12] So, you you turn it on and uh it becomes very hot and then you turn it off and becomes very cold.
[06:17] So, there was no in between.
[06:21] Um, but besides that, it was it was great.
[06:23] It was great.
[06:23] So, okay.
[06:25] So, energy consumption very important, right?
[06:27] So again what I show here is the again the the years so about a decade right and now let me tell you this is also on the logarithmic scale right seeing energy go up on the logarithmic scale not very good right
[06:40] this is a very scary prospect you know and is it is it slowing down no it doesn't seem to be slowing down so how much energy is being spent so let's put it a little bit for me it's a perspective right
[06:53] So, the average electricity used in the entire New York City area, that includes Brooklyn and Queens and and uh St. Satin Island, even though nobody ever goes there, you know, and Manhattan, um is about 134,000,000
[07:13] and Manhattan, um is about 134,000,000 megawatt hours, right?
[07:17] megawatt hours, right?
[07:17] It's it's up here so it's pretty high
[07:20] but just to train Grock 3 is already above that right just one model one
[07:23] model is already above that so this is this is very serious and that's with the
[07:26] reason that you know uh there's a lot of work right now on building data centers
[07:32] that are maybe nuclearpowered you know other other energy uh sources
[07:36] that are powered where does this end where are we going uh you know not not a
[07:38] not easy question to answer. All right.
[07:43] Um and then more more specific to this talk, what can uh photonics, right?
[07:45] We're um interested in in photonics and light, you know, what can we do on the
[07:51] energy consumption side of of the of the issue.
[07:54] Okay. So, let's look at where where is the communication, you know, in these in
[07:57] these big we we have a lot of compute, we have memory. uh where does the
[08:15] We have memory.
[08:15] Where does the communication come in, right?
[08:16] Where are the bottlenecks and and what can we do about it?
[08:18] So typically these large systems and I just go back here so you know the Gro 3 was trained on a 100,000 uh you know boxes of GPUs, right?
[08:21] Not even just not not single GPUs, these are boxes H100s.
[08:30] Now we're up to the B200s, you know the Nvidia uh you know uh box essentially.
[08:34] Um so so these are um a very very large number of compute and um when you when you have a a computing application, a workload that's spread out over so many uh so many compute elements, right?
[08:36] They need to talk to each other, there has to be communication that exists between them.
[08:38] And um there's a that's where the problem comes in.
[08:40] And so let's look at the communication.
[08:43] So if we are inside
[09:15] the communication So if we are inside what we call the socket and the socket
[09:18] what we call the socket and the socket looks like basically GPU or a small
[09:22] looks like basically GPU or a small number of GPUs right um and they're
[09:25] number of GPUs right um and they're sitting on like an interposer they're
[09:27] sitting on like an interposer they're basically you know kind of this size
[09:29] basically you know kind of this size right have some pictures I think a
[09:31] right have some pictures I think a little later on uh and so they're all
[09:35] little later on uh and so they're all integrated in this kind of 3D there's a
[09:37] integrated in this kind of 3D there's a lot of 3D integration heterogeneous
[09:39] lot of 3D integration heterogeneous integration you have your GPUs you have
[09:42] integration you have your GPUs you have your high bandwidth memory, HBMs.
[09:45] your high bandwidth memory, HBMs. Um, sometimes you might have a CPU in there,
[09:47] sometimes you might have a CPU in there, but this is kind of the basic structure.
[09:50] but this is kind of the basic structure. And so inside this this little, you
[09:53] And so inside this this little, you know, very sophisticated uh multi-chip
[09:56] know, very sophisticated uh multi-chip module, you have a lot of bandwidth.
[09:59] module, you have a lot of bandwidth. Look at these numbers. Uh, terabytes, so
[10:03] Look at these numbers. Uh, terabytes, so terabytes, not terabytes, terabytes,
[10:05] terabytes, not terabytes, terabytes, right? So 80 terabyts um or so even 100
[10:10] right? So 80 terabyts um or so even 100 terabyts is not not unusual in inside
[10:13] terabyts is not not unusual in inside this chip. But this is tremendous amount
[10:16] this chip.
[10:16] But this is tremendous amount of bandwidth.
[10:18] of bandwidth.
[10:21] However, as soon as we go out now, we have to connect them together, right?
[10:23] Remember, we have a 100,000 boxes that we have to connect to each other.
[10:27] So when we go outside of the socket, we drop by about a factor of 10 already.
[10:33] you know, we take we take a hit about a factor of 10 and then when we when we really scale the system, when we really go to the 100,000 or 20,000, whatever is the size of the system that's it's built, we take we take a huge drop by almost a factor of 100.
[10:52] So from inside the socket to across the system, there is a tremendous drop in in the communication, right?
[11:01] Okay, we you fall off a cliff.
[11:04] Why?
[11:04] Why?
[11:04] Right, that doesn't seem like it would be a good thing.
[11:09] Essentially, it's the limitation of the electrical uh communications.
[11:12] Uh to send the signal over that kind of longer distance
[11:17] over that kind of longer distance requires amplification.
[11:20] requires amplification.
[11:22] You know, the electrical signaling is is very difficult.
[11:26] Um and uh and power consumption, right?
[11:27] It takes power consumption.
[11:30] Um and also the density the density of the bandwidth is is significantly reduced.
[11:35] Density means uh how much how you know how many terabits can I put in a millimeter essentially you know the the bandwidth density.
[11:42] So all the all those things are very very difficult in in the electronic domain.
[11:46] Um and so they they just fall off a cliff.
[11:49] So what does that do to the system?
[11:54] It makes the system very very inefficient.
[11:56] So the reason that one of the reasons that you see this energy go up you know almost you know indefinitely is that of course you know the compute itself the the GPUs are very power hungry.
[12:07] They they have they have to do their work.
[12:09] The memory you know is consumes a certain amount of power but now when you we try to connect them together and they they need to efficiently work together
[12:19] they need to efficiently work together to to finish the workload to solve the
[12:22] to to finish the workload to solve the application to solve the problem.
[12:25] application to solve the problem. And so this GPU on one side of the system needs
[12:28] this GPU on one side of the system needs to access you know memory space all the
[12:31] to access you know memory space all the way on the other side of the system
[12:32] way on the other side of the system perhaps from other and so what they do
[12:34] perhaps from other and so what they do is they start from a place where they
[12:37] is they start from a place where they have a 100 times the bandwidth. They try
[12:39] have a 100 times the bandwidth. They try to maximize the amount of communication
[12:40] to maximize the amount of communication of course that they do inside the
[12:42] of course that they do inside the socket, but then they have to fetch
[12:45] socket, but then they have to fetch something some data all the way from the
[12:48] something some data all the way from the other side of the the system and then
[12:50] other side of the the system and then the bandwidth is dropping by a factor of
[12:52] the bandwidth is dropping by a factor of 100 or so. And so so while you're
[12:56] 100 or so. And so so while you're waiting for the data to come from far
[12:59] waiting for the data to come from far away, your memory space to come back
[13:01] away, your memory space to come back from far away, you're sitting there
[13:04] from far away, you're sitting there idling, wasting a lot of energy. So
[13:06] idling, wasting a lot of energy. So there's tremendous amount of extra
[13:08] there's tremendous amount of extra energy that's wasted plus you're
[13:10] energy that's wasted plus you're connecting to another box of GPUs and
[13:14] connecting to another box of GPUs and trying to get to their their memory
[13:16] trying to get to their their memory space. So it's uh you have these big
[13:20] space. So it's uh you have these big boxes that have all kinds of stuff in them, GPUs, memory, you know, sometimes FPGAAS, and there's a little imagine like a little straw of bandwidth that's connected and you're trying to put them all together, 100 thousand of them together.
[13:37] And so drinking the data out of that tiny tiny straw is is tremendously inefficient.
[13:42] Why do we why do we have systems this way?
[13:46] As I mentioned, it's the power consumption and the cost, right?
[13:50] the cost of um you know putting uh electrical connections sometimes fiber optic connections but still pluggable fiber optic connections that are you know much lower bandwidth than uh than the integrated I will talk about the integrated fatonics uh in a little bit um is is the driving the driving reason.
[14:12] So this is where the systems are today.
[14:15] Great. Um, so let me let me show you uh
[14:20] Great. Um, so let me let me show you uh an example of this exact problem and how in China they sort of try to address it.
[14:28] uh with with some interesting technologies, optical technologies.
[14:32] Okay, what what am I showing here? You might have heard on on the news that the US, you know, has some export control restrictions on the GPUs that we send to China, right? that we sell to China.
[14:47] This is not a political talk. It is what it is. Right? So in the US uh and other countries like Germany of course we have access to the best GPUs in the world which are right now there are the B200s and getting getting better right very very massive GPUs. So the B200 over here right is made by Nvidia.
[15:10] Each of the chips has uh 25,000 uh teraflops, right? Just, you know, just that that dual chip, the dual B200
[15:22] Just that that dual chip, the dual B200 has 25,000 teraflops, 2.5 pedaflops.
[15:28] I remember when I was an early stage uh researcher, we went to a whole conference where the goal of the conference was pedlop computing, right?
[15:40] We didn't think that we would ever get to a pedaflop computing.
[15:43] It was like somewhere in the future.
[15:46] Now we have it on a chip.
[15:49] Okay, this is the amazing thing.
[15:51] Okay, so so it and it is very impressive to have it.
[15:55] Um, it has a certain amount of memory.
[15:56] It has a certain amount of bandwidth.
[15:58] These are Let's put these aside for now.
[16:01] Um this the chip that the US is allowed to sell to China,
[16:07] they what they do is they they dial it down, right?
[16:09] They they they give it they give we give them a more inferior chips.
[16:15] No no no no comments on the politics but basically the chips that we sell to China is about three times less
[16:23] China is about three times less performance than what is available to the US and and other countries.
[16:26] performance than what is available to the US and and other countries.
[16:30] Okay.
[16:30] So what do we do right in the US?
[16:35] Nvidia takes their very powerful chips and they connect them with copper with electrical copper and they built uh a system.
[16:38] connect them with copper with electrical copper and they built uh a system.
[16:41] This is a small system in the scaleup domain called the NVL72.
[16:44] It can go up a little bit more.
[16:47] It's okay.
[16:47] But it's only 72 because it's limited by how far you can reach with the copper, right?
[16:53] You can't you can't scale much more than that.
[16:56] So what did what did uh China do?
[16:58] Huawei built a system called cloud matrix and they took their three times inferior chips, right?
[17:01] But they connected them with fiber optics plugg you know linear pluggable fiber optics.
[17:04] and they got two times about two times the
[17:24] two times about two times the performance of the Nvidia system with the three times better chips.
[17:31] Then they ran deepseek on it and we know we know the rest of the story.
[17:35] So everybody was talking about deepseek and the algorithm and all of that but underneath there's hardware that's using photonix.
[17:43] This is this is in my opinion the the first commercial example of a real system that's using photonix in what we call the scale up domain like actually inside the box the system connecting the GPUs together.
[17:59] Everybody uses fiber optics.
[18:01] You know, you go to data centers, you'll see lots of fiber optics, but that's in the scale out domain.
[18:05] That's to connect, you know, racks together.
[18:08] This is inside inside the system.
[18:10] And they were actually able to get better performance.
[18:11] So, this is really, really interesting.
[18:17] If you look at the papers, there are some papers on this, right?
[18:18] This is in the public domain.
[18:21] it.
[18:23] I I did searches, you know, I tried to find the word you look for the word
[18:25] to find the word you look for the word optics.
[18:28] You look for the word fatonix doesn't exist.
[18:30] doesn't exist. Very interesting.
[18:34] But, you know, um we were able to dig this out and um uh you know, it's it's in the public domain.
[18:39] I didn't I'm not showing you anything that's secret, but they're using linear pluggable optics.
[18:46] Very very interesting. Okay.
[18:50] So the other big shock wave that happened in our community um was last year actually exactly around this time.
[18:55] last year uh Nvidia made the announcement that they were going to use uh silicon phutonic integrated silicon phetonics in their systems uh inside inside the socket.
[19:10] Um, and so they showed, you know, this was a big splash and they showed that, you know, they're they're using uh everybody, it was kind of a poorly kept secret that they're fabricating their silicon photonics at TSMC.
[19:24] Uh, it's it's a mature node, right? 65 nanometers.
[19:26] mature node, right?
[19:26] 65 nanometers.
[19:26] That's a pretty mature node.
[19:29] Uh, but of course they're using TSMC 6 nanometers for the for the electric.
[19:34] This is in Taiwan.
[19:34] Yeah.
[19:34] Yeah.
[19:34] Yeah.
[19:36] Absolutely.
[19:36] They they make Nvidia makes everything in TSMC.
[19:38] Yeah.
[19:38] Yeah.
[19:40] And so does AMD and so does Broadcom.
[19:42] Basically TSMC is the the go-to.
[19:42] Yeah.
[19:45] But now TSMC is doing fatonix.
[19:45] That's they've been working on it for a couple years, but now it's it's really happening.
[19:51] Yeah.
[19:51] And they're using an old node for photonics.
[19:55] The good thing about fatonics is that you don't need an advanced node, right?
[20:00] Because what we care the fatonic wavelength is that we're using here typically it's either 1.3 microns 1.5 microns it's huge right so the only thing you care about in terms of the the cos process is that you have enough um uh that that you have enough ability to to um print the wave guides and the other photonic devices the the resonators so that you don't have scattering right so you have low losses
[20:29] scattering right so you have low losses the other The other thing though is you
[20:32] the other The other thing though is you want to have the advanced tools right so
[20:36] want to have the advanced tools right so a foundry of course it's it's it's all
[20:38] a foundry of course it's it's it's all about the node right how precise you
[20:40] about the node right how precise you know two nanometers angstroms now you
[20:43] know two nanometers angstroms now you know people are going further and
[20:44] know people are going further and further for optics we don't care so much
[20:47] further for optics we don't care so much about that right 65 nanometers is plenty
[20:50] about that right 65 nanometers is plenty right however when you have uh very
[20:55] right however when you have uh very advanced uh foundry tools then you can
[20:59] advanced uh foundry tools then you can also make them with very high um uh very
[21:03] also make them with very high um uh very low I should say variations right so the
[21:08] low I should say variations right so the old foundaries the one that are using
[21:10] old foundaries the one that are using you know 200 mm wafers and 180 nanometer
[21:15] you know 200 mm wafers and 180 nanometer nodes or whatever they also have old
[21:17] nodes or whatever they also have old tools and those tools are not very
[21:20] tools and those tools are not very precise right the very we we care about
[21:23] precise right the very we we care about the variations we want the waveguides
[21:26] the variations we want the waveguides the rings the manders that we make in
[21:28] the rings the manders that we make in silicon photonics to be exactly the same
[21:31] silicon photonics to be exactly the same over and over again and you know we know
[21:33] over and over again and you know we know in optics you know that's that's uh
[21:35] in optics you know that's that's uh requires that's where the precision is
[21:37] requires that's where the precision is important. Um and so the tool it's about
[21:39] important. Um and so the tool it's about the tools it's about the way for scale
[21:41] the tools it's about the way for scale tools that actually matters. So anyways,
[21:43] tools that actually matters. So anyways, this was a big announcement.
[21:46] this was a big announcement. It was made uh it was actually made in
[21:48] It was made uh it was actually made in March of 2025 right before the OFC. If
[21:52] March of 2025 right before the OFC. If you guys are familiar, that's the big
[21:54] you guys are familiar, that's the big conference for us. OFC 2025. So
[21:57] conference for us. OFC 2025. So everybody at OFC was crazy about this.
[22:00] everybody at OFC was crazy about this. Um and what they did was they said we
[22:03] Um and what they did was they said we are doing this right until then
[22:05] are doing this right until then everybody was talking about it. But when
[22:07] everybody was talking about it. But when Nvidia said we're doing this, that
[22:09] Nvidia said we're doing this, that really changed changed the the game.
[22:12] really changed changed the the game. And very importantly, here's what
[22:14] And very importantly, here's what they're doing. They're not doing
[22:16] they're doing. They're not doing manders.
[22:18] manders. They're doing rings. They're doing
[22:20] They're doing rings. They're doing rings. Why are they doing rings? You
[22:23] rings. Why are they doing rings? You know, rings are they're temperature
[22:25] know, rings are they're temperature sensitive. You know, they're wavelength
[22:28] sensitive. You know, they're wavelength sensitive. You just put mock zenders
[22:30] sensitive. You just put mock zenders there. They're much better. They're
[22:33] there. They're much better. They're doing rings because this is the only way
[22:36] doing rings because this is the only way to get the bandwidth density and the low
[22:38] to get the bandwidth density and the low energy consumption inside the socket. If
[22:42] energy consumption inside the socket. If I use MCIs, if I use, you know, any
[22:45] I use MCIs, if I use, you know, any other technology really, direct drive
[22:47] other technology really, direct drive lasers, you know, all these other
[22:49] lasers, you know, all these other things, they're going to be too large uh
[22:52] things, they're going to be too large uh and and too power consuming to get to
[22:55] and and too power consuming to get to get to the the um the metrics that I
[22:58] get to the the um the metrics that I need. Of course, they only showed one
[23:00] need. Of course, they only showed one ring. In in a more recent work um uh
[23:03] ring. In in a more recent work um uh just now that came out in in in ISSCC,
[23:06] just now that came out in in in ISSCC, another another conference, electronics
[23:08] another another conference, electronics conference, they showed eight eight
[23:10] conference, they showed eight eight channels. So, so this is really this is
[23:13] channels. So, so this is really this is going very quickly, very quickly. Now,
[23:16] going very quickly, very quickly. Now, there's lots of uh startup companies
[23:18] there's lots of uh startup companies that are doing this. If you've heard of
[23:21] that are doing this. If you've heard of um celestial AI uh uh light matter, I
[23:26] um celestial AI uh uh light matter, I have a company as well that that's
[23:27] have a company as well that that's working on this. So this is a very very
[23:30] working on this. So this is a very very exciting field right now. Lots of lots
[23:32] exciting field right now. Lots of lots of action.
[23:35] of action. What they also need that's missing is
[23:38] What they also need that's missing is the lasers. Sometimes you know people in
[23:40] the lasers. Sometimes you know people in this field you know they they came at it
[23:42] this field you know they they came at it from different areas you know they're
[23:43] from different areas you know they're from computing from electronics
[23:46] from computing from electronics and and they're like oh silicon
[23:48] and and they're like oh silicon photonics we can integrate everything on
[23:50] photonics we can integrate everything on a chip it's beautiful it'll be just like
[23:52] a chip it's beautiful it'll be just like CMOS it's great oh we need a laser they
[23:58] CMOS it's great oh we need a laser they forget that to do photonics you need
[24:00] forget that to do photonics you need lasers I I know I know it sounds
[24:03] lasers I I know I know it sounds ridiculous the Mox plunk for light but
[24:06] ridiculous the Mox plunk for light but to do light you need lasers I mean in in
[24:08] to do light you need lasers I mean in in this way, the integrated form. So, of
[24:11] this way, the integrated form. So, of course, you know, I'm being I'm being,
[24:14] course, you know, I'm being I'm being, you know, silly, but but it's it's a bit
[24:17] you know, silly, but but it's it's a bit of an afterthought, right? Um, and so
[24:21] of an afterthought, right? Um, and so they made, you know, Nvidia made this
[24:22] they made, you know, Nvidia made this big splash, but where's the laser? So
[24:26] big splash, but where's the laser? So they have some kind of box that's
[24:28] they have some kind of box that's sitting somewhere off the system and
[24:30] sitting somewhere off the system and they really need not just a single
[24:33] they really need not just a single wavelength laser, they really need uh
[24:36] wavelength laser, they really need uh multiple wavelength lasers, dense WDM
[24:38] multiple wavelength lasers, dense WDM lasers to make this happen. And um if
[24:42] lasers to make this happen. And um if you there's a story if you just go to
[24:46] you there's a story if you just go to NASDAQ, right, stock market, look at a
[24:49] NASDAQ, right, stock market, look at a company that you might know called
[24:51] company that you might know called momentum or coherent. Everybody knows
[24:54] momentum or coherent. Everybody knows Coherent, right? They make lasers. Look
[24:56] Coherent, right? They make lasers. Look at their stock price in the last few
[24:59] at their stock price in the last few months.
[25:01] months. Over the last year, Lmentum went, I
[25:04] Over the last year, Lmentum went, I think, from 1 billion to 50 billion.
[25:06] think, from 1 billion to 50 billion. What's going on here? Everybody's
[25:09] What's going on here? Everybody's realizing we need lasers. There aren't
[25:12] realizing we need lasers. There aren't many lasers in the world.
[25:14] many lasers in the world. I'm not giving any investment advice by
[25:17] I'm not giving any investment advice by the way but you know
[25:21] the way but you know you guys are optics experts so you can
[25:23] you guys are optics experts so you can you can think of okay so how how do we
[25:25] you can think of okay so how how do we build these things how do they work
[25:29] build these things how do they work our approach and I think it's uh you'll
[25:31] our approach and I think it's uh you'll see why we we're very excited about it
[25:33] see why we we're very excited about it is to use uh comb laser technology I
[25:36] is to use uh comb laser technology I know that uh my colleague uh professor
[25:38] know that uh my colleague uh professor Mkha Lipson was here just just uh last
[25:41] Mkha Lipson was here just just uh last week maybe and uh so we work together on
[25:43] week maybe and uh so we work together on this with uh Lipson Gayeda and and
[25:46] this with uh Lipson Gayeda and and myself on developing these comb laser
[25:49] myself on developing these comb laser technology. So this is why why is this
[25:51] technology. So this is why why is this so interesting? Because we have a single
[25:53] so interesting? Because we have a single we can use a single laser. We still need
[25:55] we can use a single laser. We still need a laser but we can generate lots and
[25:57] a laser but we can generate lots and lots of different channels lots of lots
[25:59] lots of different channels lots of lots and lots of different wavelength
[26:00] and lots of different wavelength channels all at the same time that are
[26:03] channels all at the same time that are precisely aligned for our data. And so
[26:06] precisely aligned for our data. And so this is the t the typical architecture.
[26:09] this is the t the typical architecture. Imagine that we have, you know, a comb
[26:11] Imagine that we have, you know, a comb that's generating all these channels and
[26:13] that's generating all these channels and then we go into our silicon photonic
[26:15] then we go into our silicon photonic chip. We use um we use these resonator
[26:19] chip. We use um we use these resonator uh modulators. So they do two things at
[26:22] uh modulators. So they do two things at once. They they are resonators. So they
[26:25] once. They they are resonators. So they resonate at one specific they're
[26:27] resonate at one specific they're designed to be resonant at one specific
[26:30] designed to be resonant at one specific wavelength which means that they are
[26:33] wavelength which means that they are selected they select that particular
[26:35] selected they select that particular wavelength. Right? So I don't need to go
[26:37] wavelength. Right? So I don't need to go if I if I use the Marander and I know
[26:40] if I if I use the Marander and I know it's very simple but just just bear with
[26:42] it's very simple but just just bear with me. If I had if I had Marander
[26:44] me. If I had if I had Marander modulators and I come in with multiple
[26:47] modulators and I come in with multiple wavelengths, right? I need to first
[26:51] wavelengths, right? I need to first separate out those wavelengths with some
[26:53] separate out those wavelengths with some kind of a wavelength demiplexer, right?
[26:57] kind of a wavelength demiplexer, right? And why is that not good? It's not good
[27:00] And why is that not good? It's not good because that those WDM demultiplexers
[27:03] because that those WDM demultiplexers are huge. you know, it's hard to make
[27:05] are huge. you know, it's hard to make them very small. Um, and then and then
[27:08] them very small. Um, and then and then after that, I put these huge humongous,
[27:11] after that, I put these huge humongous, you know, MZIs. The MZIs are like, you
[27:13] you know, MZIs. The MZIs are like, you know, we we size MZIs in hundreds of
[27:16] know, we we size MZIs in hundreds of microns or even submillimeter, right?
[27:19] microns or even submillimeter, right? That's like, forget it. The whole size
[27:21] That's like, forget it. The whole size of the chip is a millimeter, right? So,
[27:23] of the chip is a millimeter, right? So, we can't use that stuff for because
[27:26] we can't use that stuff for because otherwise if we look at the bandwidth
[27:27] otherwise if we look at the bandwidth density that we generate, right? It's
[27:30] density that we generate, right? It's not just the bandwidth. It's not just
[27:31] not just the bandwidth. It's not just that we can run the Mander at 100 Gbit.
[27:34] that we can run the Mander at 100 Gbit. It's about the bandwidth density. How
[27:36] It's about the bandwidth density. How much can I get in and out of the chip?
[27:39] much can I get in and out of the chip? So, so having a device that's a
[27:41] So, so having a device that's a resonator
[27:43] resonator does two things at once. It does the
[27:45] does two things at once. It does the modulation and this the wavelength
[27:47] modulation and this the wavelength selection is very advantageous. So, very
[27:49] selection is very advantageous. So, very important. So, we use these uh
[27:51] important. So, we use these uh modulators and then we have the opposite
[27:53] modulators and then we have the opposite on the opposite side. We just have
[27:55] on the opposite side. We just have filters and detectors. And this is kind
[27:57] filters and detectors. And this is kind of the basic structure. And then you
[27:59] of the basic structure. And then you know it you'll you'll see that there's
[28:01] know it you'll you'll see that there's um
[28:05] >> it's basically by shifting the
[28:06] >> it's basically by shifting the resonance. Yes, there's there definitely
[28:08] resonance. Yes, there's there definitely there going to be losses but it's not
[28:11] there going to be losses but it's not like an electro electro absorption
[28:12] like an electro electro absorption modulator. It doesn't it doesn't work on
[28:14] modulator. It doesn't it doesn't work on on that loss. Um it's it's you shift the
[28:18] on that loss. Um it's it's you shift the resonance and that that becomes the the
[28:20] resonance and that that becomes the the the onoff. Yes. So there's some some
[28:22] the onoff. Yes. So there's some some tweaks there. What we do is so there are
[28:26] tweaks there. What we do is so there are various designs on the resonator. Um you
[28:29] various designs on the resonator. Um you can have a ring. What we do uh and I
[28:31] can have a ring. What we do uh and I find out that this is also down here. We
[28:34] find out that this is also down here. We we use discs and discs um instead of the
[28:37] we use discs and discs um instead of the the ring which propagates the light you
[28:40] the ring which propagates the light you know along the waveguide the the disk is
[28:43] know along the waveguide the the disk is it's a whispering gallery mode that we
[28:45] it's a whispering gallery mode that we design. Why do we use disk? I'll go into
[28:47] design. Why do we use disk? I'll go into it a little bit but primarily because
[28:49] it a little bit but primarily because you can make it smaller, right? We're
[28:51] you can make it smaller, right? We're trying to make things optics is way too
[28:53] trying to make things optics is way too big. I'm sorry. it's just, you know, the
[28:55] big. I'm sorry. it's just, you know, the wavelength is just too big when you're
[28:57] wavelength is just too big when you're thinking about integrating it with
[28:58] thinking about integrating it with electronics. And so everything that we
[29:01] electronics. And so everything that we can do to squeeze the the size is very
[29:03] can do to squeeze the the size is very is very important. So this is kind of
[29:06] is very important. So this is kind of the results.
[29:08] the results. You can go to sleep after this if you
[29:10] You can go to sleep after this if you want. Um, but the results are
[29:14] want. Um, but the results are uh here's here's the most important
[29:17] uh here's here's the most important metric that we care about, right? I keep
[29:20] metric that we care about, right? I keep going back to it. It's about what is
[29:22] going back to it. It's about what is your bandwidth density, right? How much
[29:25] your bandwidth density, right? How much bandwidth do you have per unit
[29:26] bandwidth do you have per unit millimeter or per unit millimeter
[29:28] millimeter or per unit millimeter squared and how much energy per bit are
[29:31] squared and how much energy per bit are you spending? Those are the two things.
[29:34] you spending? Those are the two things. And the key is you need both at the same
[29:36] And the key is you need both at the same time. If I just wanted bandwidth density
[29:40] time. If I just wanted bandwidth density at any energy consumption, I could
[29:42] at any energy consumption, I could probably figure out a way to do that
[29:43] probably figure out a way to do that with electronics, right? you know I I
[29:47] with electronics, right? you know I I can electronics on on on a chip can be
[29:50] can electronics on on on a chip can be very very good in terms of the bandwidth
[29:52] very very good in terms of the bandwidth density. I can put lots and lots of
[29:53] density. I can put lots and lots of interconnects, you know, multiple
[29:55] interconnects, you know, multiple layers, but I need to be able to move
[29:59] layers, but I need to be able to move that data across the system. Um, and so,
[30:04] that data across the system. Um, and so, you know, doing it with electronics is
[30:06] you know, doing it with electronics is not going to work well. And, um, let me
[30:09] not going to work well. And, um, let me give you another example. Let's say the
[30:10] give you another example. Let's say the pluggable optics, right? Everybody knows
[30:12] pluggable optics, right? Everybody knows the pluggable optics. You know, you have
[30:14] the pluggable optics. You know, you have a transceiver. If you go to a
[30:17] a transceiver. If you go to a conference, you know, they're going to
[30:18] conference, you know, they're going to talk about, oh, we can do 800 gigabit
[30:21] talk about, oh, we can do 800 gigabit per second, you know, in pluggable
[30:23] per second, you know, in pluggable optics, maybe soon, you know, 1.6
[30:26] optics, maybe soon, you know, 1.6 terabit. Not enough because the
[30:29] terabit. Not enough because the pluggable optic is like huge. It's like,
[30:31] pluggable optic is like huge. It's like, you know, it's like the size the size of
[30:32] you know, it's like the size the size of a finger or so. And you're like, oh,
[30:35] a finger or so. And you're like, oh, that's very nice. I can get a terabit
[30:37] that's very nice. I can get a terabit from this. No, not enough. We need tens
[30:39] from this. No, not enough. We need tens of terabits per millimeter. That's the
[30:42] of terabits per millimeter. That's the bandwidth density that we need to get.
[30:44] bandwidth density that we need to get. um and at the same time very low energy
[30:46] um and at the same time very low energy consumption. So the way the only way we
[30:49] consumption. So the way the only way we can really get there is by increasing
[30:51] can really get there is by increasing the number of wavelengths. If we have a
[30:54] the number of wavelengths. If we have a lot of wavelengths, we can modulate each
[30:57] lot of wavelengths, we can modulate each wavelength at a relatively modest data
[30:59] wavelength at a relatively modest data rate that reduces the energy consumption
[31:02] rate that reduces the energy consumption and we can if we can squeeze enough of
[31:04] and we can if we can squeeze enough of this on a chip, we can get the bandwidth
[31:06] this on a chip, we can get the bandwidth density. So again what you see here and
[31:08] density. So again what you see here and I want to say this is again a
[31:10] I want to say this is again a logarithmic scale. So here here is some
[31:13] logarithmic scale. So here here is some different uh examples depending on the
[31:16] different uh examples depending on the number of of channels that you you use.
[31:19] number of of channels that you you use. The greener it is the less energy
[31:21] The greener it is the less energy consumption. The more green it is right
[31:23] consumption. The more green it is right the less energy consumption. So so this
[31:26] the less energy consumption. So so this is this is some of our our recent work.
[31:28] is this is some of our our recent work. You know we can we can get very very low
[31:30] You know we can we can get very very low energy consumption and we do it by using
[31:33] energy consumption and we do it by using a lot of wavelengths 80 wavelengths in
[31:35] a lot of wavelengths 80 wavelengths in this case. Um this is the Nvidia. It's
[31:38] this case. Um this is the Nvidia. It's not bad. It's not bad. They they you
[31:40] not bad. It's not bad. They they you know they they run their their
[31:42] know they they run their their modulators at at at pretty high speeds
[31:45] modulators at at at pretty high speeds and so they're not as energy efficient
[31:47] and so they're not as energy efficient but they're honestly they're not using a
[31:50] but they're honestly they're not using a lot of wavelengths because they don't
[31:51] lot of wavelengths because they don't have a lot of wavelengths, right? There
[31:53] have a lot of wavelengths, right? There isn't available you know you can't buy
[31:57] isn't available you know you can't buy 128 laser source right now in large
[32:00] 128 laser source right now in large scale. So that's why they're they're
[32:02] scale. So that's why they're they're over here. They would love to be here
[32:04] over here. They would love to be here where the research is but they're not
[32:06] where the research is but they're not they can't get there.
[32:08] they can't get there. >> Yeah. This is the one dimensional
[32:11] >> Yeah. This is the one dimensional cross-section in the surface of the uh
[32:14] cross-section in the surface of the uh >> it's it's uh that exactly it's the
[32:17] >> it's it's uh that exactly it's the perimeter the perimeter. So this is the
[32:20] perimeter the perimeter. So this is the most precious uh real estate on the chip
[32:23] most precious uh real estate on the chip because on the chip you're going to have
[32:24] because on the chip you're going to have a lot of uh electronics that's very very
[32:27] a lot of uh electronics that's very very important right the GPUs the other stuff
[32:30] important right the GPUs the other stuff and then you have a little bit you try
[32:31] and then you have a little bit you try to minimize how much you you have for
[32:34] to minimize how much you you have for your IO uh you you're not going to use
[32:36] your IO uh you you're not going to use optics inside the chip it doesn't make
[32:39] optics inside the chip it doesn't make any sense because electronics is just
[32:40] any sense because electronics is just too good electronics is really good um
[32:44] too good electronics is really good um so it's about moving the data out of the
[32:46] so it's about moving the data out of the chip um chip to chip or across the
[32:49] chip um chip to chip or across the system and so the perimeter is the most
[32:51] system and so the perimeter is the most important parameter and that's the most
[32:53] important parameter and that's the most precious real estate. Yeah. Exactly.
[32:56] precious real estate. Yeah. Exactly. Exactly. There are other you know other
[32:58] Exactly. There are other you know other ways like let's say you want to come
[32:59] ways like let's say you want to come from the top but you know that's that's
[33:02] from the top but you know that's that's there's always a trade-off because
[33:03] there's always a trade-off because people want to put your your memory
[33:05] people want to put your your memory integrated on the top heterogeneous
[33:07] integrated on the top heterogeneous integrated memory because the memory
[33:09] integrated memory because the memory bandwidth is really key. So it's a
[33:12] bandwidth is really key. So it's a combination of understanding the
[33:14] combination of understanding the packaging what's the packaging approach
[33:16] packaging what's the packaging approach and and where you can squeeze in your
[33:18] and and where you can squeeze in your your IO. Yeah. So sometimes we talk
[33:21] your IO. Yeah. So sometimes we talk about millimeter squared but usually the
[33:24] about millimeter squared but usually the millimeter is is another is important
[33:25] millimeter is is another is important parameter.
[33:27] parameter. Okay. Actually that leads to this
[33:29] Okay. Actually that leads to this picture. So this is kind of where things
[33:32] picture. So this is kind of where things are today. This is the reason pluggable
[33:34] are today. This is the reason pluggable optics it's very good. It's very useful.
[33:37] optics it's very good. It's very useful. It's going to be continued.
[33:39] It's going to be continued. Well, I'm not saying that pluggable
[33:40] Well, I'm not saying that pluggable optics is going out of out of uh
[33:43] optics is going out of out of uh fashion. It continues. Don't worry about
[33:44] fashion. It continues. Don't worry about it. It's all good. But the new and this
[33:49] it. It's all good. But the new and this is what this uh socket I kept kept
[33:51] is what this uh socket I kept kept talking about the socket. This is sort
[33:52] talking about the socket. This is sort of what it looks like, right? You have
[33:54] of what it looks like, right? You have some GPUs, you have your memory. You can
[33:56] some GPUs, you have your memory. You can see this is very crowded, right? This is
[33:59] see this is very crowded, right? This is kind of the TSMC, you know, 3D packaging
[34:03] kind of the TSMC, you know, 3D packaging approach that they have. Me, many
[34:05] approach that they have. Me, many different things around this. Um and the
[34:09] different things around this. Um and the idea is that can we bring the photonics
[34:11] idea is that can we bring the photonics in some way you know inside inside this
[34:14] in some way you know inside inside this package and uh you know should it be 3D
[34:18] package and uh you know should it be 3D that's that's what we like to do so that
[34:20] that's that's what we like to do so that you can get a lot of density uh from
[34:22] you can get a lot of density uh from that
[34:24] that but this is still ongoing ongoing work
[34:27] but this is still ongoing ongoing work uh you know all right so let's go back
[34:30] uh you know all right so let's go back to our story about Nvidia and Huawei um
[34:34] to our story about Nvidia and Huawei um what if we replaced
[34:37] what if we replaced We we had we took the best of
[34:39] We we had we took the best of everything, right? Take the best GPUs in
[34:42] everything, right? Take the best GPUs in the world. But now instead of using the
[34:44] the world. But now instead of using the pluggable optics that Huawei had to use
[34:47] pluggable optics that Huawei had to use because that that's what's available
[34:48] because that that's what's available commercially, we use this new optics
[34:52] commercially, we use this new optics um that's that's available. I I like to
[34:55] um that's that's available. I I like to call it embedded futonics, right?
[34:56] call it embedded futonics, right? Because we're really embedding it inside
[34:58] Because we're really embedding it inside the socket. Look at these numbers. You
[35:01] the socket. Look at these numbers. You know, we can I mean the scaling is just
[35:04] know, we can I mean the scaling is just amazing. And this is all under the same
[35:06] amazing. And this is all under the same energy envelope that we started from. So
[35:09] energy envelope that we started from. So we're not spending any more energy.
[35:11] we're not spending any more energy. We're just vastly increasing the amount
[35:14] We're just vastly increasing the amount of of of uh bandwidth that's available
[35:17] of of of uh bandwidth that's available to the system. So now we can basically
[35:21] to the system. So now we can basically cut out that 100x
[35:24] cut out that 100x drop that we used to have that we have
[35:27] drop that we used to have that we have today in the current systems, right? We
[35:28] today in the current systems, right? We went from the socket, you know, to the
[35:30] went from the socket, you know, to the to the system. We took a 100 100x uh
[35:33] to the system. We took a 100 100x uh drop.
[35:34] drop. We can eliminate it. And I like to say
[35:37] We can eliminate it. And I like to say the world is my cash. I know it's it's a
[35:39] the world is my cash. I know it's it's a CS kind of joke, right? So I can
[35:42] CS kind of joke, right? So I can communicate anywhere in the system like
[35:44] communicate anywhere in the system like it was my own personal cash. Um I wish
[35:48] it was my own personal cash. Um I wish it was cash, you know, with the dollar
[35:49] it was cash, you know, with the dollar signs, but it's okay.
[35:52] signs, but it's okay. All right, here we go. So you're
[35:54] All right, here we go. So you're probably very familiar with the world of
[35:56] probably very familiar with the world of silicon photonic fabrication there. So
[35:58] silicon photonic fabrication there. So today, you know, a lot of the big
[36:00] today, you know, a lot of the big foundaries, the commercial foundaries
[36:02] foundaries, the commercial foundaries that are doing electronics, I mentioned
[36:04] that are doing electronics, I mentioned TSMC,
[36:05] TSMC, um are now doing photonics. Global
[36:07] um are now doing photonics. Global foundaries is doing photonics. Tower is
[36:10] foundaries is doing photonics. Tower is doing photonics and and and others
[36:12] doing photonics and and and others others across the world. Um AIM is a is
[36:16] others across the world. Um AIM is a is a is is is it's used by IBM in in
[36:20] a is is is it's used by IBM in in Albony, New York. Um so it's it's
[36:22] Albony, New York. Um so it's it's basically it does 300 millimeter, right?
[36:26] basically it does 300 millimeter, right? 12-in wafer photonix. So big time
[36:29] 12-in wafer photonix. So big time phetonics, but it's it's a research kind
[36:32] phetonics, but it's it's a research kind of like a re R&D institute. It's not a
[36:35] of like a re R&D institute. It's not a commercial entity. So it has the best of
[36:39] commercial entity. So it has the best of the worlds that we care about which is
[36:41] the worlds that we care about which is we can design new things with them. If
[36:44] we can design new things with them. If if I work with TSMC or if I work with
[36:47] if I work with TSMC or if I work with global funeries, which we do, I can't
[36:49] global funeries, which we do, I can't design I can't tell TSMC, oh no, change
[36:51] design I can't tell TSMC, oh no, change change your process, you know, add
[36:53] change your process, you know, add doping, you know, they're not going to
[36:55] doping, you know, they're not going to listen to me. Um
[36:57] listen to me. Um uh but I can work with AIM to do that in
[37:01] uh but I can work with AIM to do that in a in in a more reasonable, you know,
[37:04] a in in a more reasonable, you know, costwise. So, but I have access to the
[37:07] costwise. So, but I have access to the 300 millimeter FAB. So, this is the
[37:09] 300 millimeter FAB. So, this is the night really really nice thing. And so
[37:11] night really really nice thing. And so this aim has been around since about uh
[37:14] this aim has been around since about uh 2015. Uh it's it's a manufacturing
[37:17] 2015. Uh it's it's a manufacturing institute that was sponsored by
[37:19] institute that was sponsored by primarily by the government in the US
[37:21] primarily by the government in the US and I've been involved with it since
[37:23] and I've been involved with it since since the beginning. So it's been a
[37:24] since the beginning. So it's been a great uh a great collaboration and so
[37:28] great uh a great collaboration and so usually uh when when you design a chip
[37:31] usually uh when when you design a chip any chip whether it's CMOS or Phetonics
[37:35] any chip whether it's CMOS or Phetonics uh you you do something as a university
[37:37] uh you you do something as a university person you do something called MPW right
[37:39] person you do something called MPW right multi-RO wafer run which means that you
[37:42] multi-RO wafer run which means that you buy little piece of the real estate on
[37:44] buy little piece of the real estate on the wafer you know I want to buy 5
[37:46] the wafer you know I want to buy 5 millime squared or 25 millime squared
[37:49] millime squared or 25 millime squared whatever it is you pay a lot of money
[37:51] whatever it is you pay a lot of money you put some designs and and six months
[37:54] you put some designs and and six months later or five months later, you get your
[37:56] later or five months later, you get your chips back, they don't work, and then
[37:58] chips back, they don't work, and then you realize that you just spent a lot of
[37:59] you realize that you just spent a lot of money on a little piece of silicon.
[38:02] money on a little piece of silicon. It happens, right? Uh so this this is we
[38:05] It happens, right? Uh so this this is we do that all the time. What we're able to
[38:08] do that all the time. What we're able to do now is instead of just buying a
[38:11] do now is instead of just buying a little piece, we get the entire reticle,
[38:14] little piece, we get the entire reticle, right? The entire 300 millimeter
[38:16] right? The entire 300 millimeter reticle. And so we can custom design
[38:19] reticle. And so we can custom design everything. Um, and so we've been doing
[38:21] everything. Um, and so we've been doing this for about a decade now, a little
[38:23] this for about a decade now, a little bit less than a decade. Um, and uh, and
[38:26] bit less than a decade. Um, and uh, and so this was our very first full reticle.
[38:30] so this was our very first full reticle. And so you can see it has I mean it's
[38:32] And so you can see it has I mean it's it's amazing. There's almost too much
[38:34] it's amazing. There's almost too much space, right? The students get very
[38:36] space, right? The students get very spoiled. You just they can draw all
[38:38] spoiled. You just they can draw all kinds of things. It also enables you to
[38:40] kinds of things. It also enables you to do a lot of statistical work. Um, so you
[38:44] do a lot of statistical work. Um, so you can see what the fabrication variations
[38:46] can see what the fabrication variations are across a wafer and all kinds of
[38:48] are across a wafer and all kinds of stuff. It's just and the and lastly the
[38:51] stuff. It's just and the and lastly the reason it's very interesting is that it
[38:54] reason it's very interesting is that it allows you to do wafer scale packaging
[38:57] allows you to do wafer scale packaging right if I just have my little die right
[38:59] right if I just have my little die right that I got back from my MPW run now I
[39:02] that I got back from my MPW run now I want to say well let let me put an
[39:04] want to say well let let me put an electronic chip on top of it to do some
[39:06] electronic chip on top of it to do some 3D integration and I I go to a place
[39:09] 3D integration and I I go to a place that does packaging and I give them my
[39:11] that does packaging and I give them my die and they're like we can't do
[39:14] die and they're like we can't do anything with this you know we can't
[39:15] anything with this you know we can't align it you know we can't if you give
[39:17] align it you know we can't if you give me a wafer
[39:19] me a wafer then then I can use my 300 millimeter
[39:21] then then I can use my 300 millimeter tools. That's why the tools are very
[39:23] tools. That's why the tools are very very important. It's all about the
[39:24] very important. It's all about the tools. Uh the TSMC even though they have
[39:27] tools. Uh the TSMC even though they have a 65 nanometer process, this is also a
[39:30] a 65 nanometer process, this is also a 65 nanometer process. It's a noble
[39:32] 65 nanometer process. It's a noble process, but the tools 300 millimeter
[39:35] process, but the tools 300 millimeter tools are the key to to uh everything.
[39:39] tools are the key to to uh everything. So this was our first wafer. Then we got
[39:41] So this was our first wafer. Then we got greedy. We did another wafer and another
[39:44] greedy. We did another wafer and another wafer. And this this one is still in
[39:47] wafer. And this this one is still in fabrication. So it's very very nice um
[39:50] fabrication. So it's very very nice um to to so so to do this kind of work
[39:53] to to so so to do this kind of work right this kind of very complex
[39:55] right this kind of very complex integrated photonic circuitry work
[39:58] integrated photonic circuitry work having this capability is is is
[40:00] having this capability is is is gamechanging for us
[40:03] gamechanging for us and uh this is just an example you're
[40:05] and uh this is just an example you're right right how do you how do you
[40:07] right right how do you how do you measure all these things right you have
[40:09] measure all these things right you have a wafer you know um you go to a
[40:13] a wafer you know um you go to a conventional lab and you know you have
[40:15] conventional lab and you know you have little microscope and probes and
[40:17] little microscope and probes and everybody body's measuring their chips.
[40:19] everybody body's measuring their chips. Okay, there's, you know, if you need to
[40:21] Okay, there's, you know, if you need to measure a wafer, especially to get
[40:23] measure a wafer, especially to get statistical data, forget about it,
[40:25] statistical data, forget about it, right? It's going to take, you know,
[40:26] right? It's going to take, you know, each PhD student a long time. So we were
[40:30] each PhD student a long time. So we were able to procure this uh this is a ficon
[40:33] able to procure this uh this is a ficon tech probber but it's it's very
[40:35] tech probber but it's it's very specifically designed for us where we
[40:37] specifically designed for us where we can do RF DC and optical both optical uh
[40:41] can do RF DC and optical both optical uh vertical as well as optical um uh uh you
[40:46] vertical as well as optical um uh uh you know to uh directly to the chip coupling
[40:48] know to uh directly to the chip coupling automated everything is automated so now
[40:51] automated everything is automated so now we can get wafer scale data it it's not
[40:54] we can get wafer scale data it it's not as magical as I'm saying you know
[40:56] as magical as I'm saying you know there's a lot of tweaking and things so
[40:58] there's a lot of tweaking and things so there's a couple the PhDs that work with
[41:00] there's a couple the PhDs that work with this. It's not but it's a lot more
[41:02] this. It's not but it's a lot more enabling than than we could. All right.
[41:05] enabling than than we could. All right. Um Oh, so the comb, right? The comb is a
[41:09] Um Oh, so the comb, right? The comb is a really key thing. And uh so as I
[41:12] really key thing. And uh so as I mentioned, this is primarily led by Geda
[41:14] mentioned, this is primarily led by Geda and Lipson.
[41:16] and Lipson. This is a this is a picture of just in
[41:19] This is a this is a picture of just in our lab of of the comb. It's a very
[41:22] our lab of of the comb. It's a very simple chip. It's basically a silicon
[41:24] simple chip. It's basically a silicon nitride resonator chip that's pumped by
[41:27] nitride resonator chip that's pumped by laser. So you still need a laser. Uh but
[41:30] laser. So you still need a laser. Uh but the the nice thing about it is that the
[41:32] the the nice thing about it is that the pump can sit anywhere. It just comes in
[41:35] pump can sit anywhere. It just comes in with a fiber. So it can sit anywhere in
[41:37] with a fiber. So it can sit anywhere in the system. And then this this little uh
[41:40] the system. And then this this little uh silicon nitride chip is you know just
[41:42] silicon nitride chip is you know just has a little thermal control on it. Very
[41:44] has a little thermal control on it. Very very simple. You can we are now working
[41:47] very simple. You can we are now working on a new project where the the the comb
[41:52] on a new project where the the the comb resonator and the active silicon fatonic
[41:54] resonator and the active silicon fatonic will all be on the same chip. So right
[41:57] will all be on the same chip. So right now it's two separate chips. We have the
[41:59] now it's two separate chips. We have the comb and then we have the link. But
[42:01] comb and then we have the link. But because silicon nitrite is part of the
[42:03] because silicon nitrite is part of the process, part of the the CMOS process,
[42:06] process, part of the the CMOS process, it just needs to be a little thick. Um
[42:08] it just needs to be a little thick. Um but that that's what we're working on
[42:09] but that that's what we're working on right now. So these are some actually
[42:11] right now. So these are some actually some new results um from that we were
[42:15] some new results um from that we were able to to get from from the combs. So
[42:18] able to to get from from the combs. So typically these combs run at a certain
[42:21] typically these combs run at a certain uh FSR, right? Free spectral range. And
[42:24] uh FSR, right? Free spectral range. And so this particular comb was designed to
[42:27] so this particular comb was designed to have an FSR of 100 gigahertz. But it
[42:30] have an FSR of 100 gigahertz. But it turns out that
[42:32] turns out that um this is from uh professor Geda's Alex
[42:35] um this is from uh professor Geda's Alex Geda's group. They were able to run at
[42:38] Geda's group. They were able to run at different at different uh FSR states and
[42:41] different at different uh FSR states and so they can go at 200 gigahertz and 300
[42:45] so they can go at 200 gigahertz and 300 gigahertz. And this is just beautiful.
[42:48] gigahertz. And this is just beautiful. we have flexibility in the FSR and it
[42:51] we have flexibility in the FSR and it really helps on the design of of of the
[42:53] really helps on the design of of of the link. So just want to mention that and
[42:56] link. So just want to mention that and it's uh you know pretty very high power
[42:59] it's uh you know pretty very high power as well uh for for combs. Okay. The the
[43:03] as well uh for for combs. Okay. The the modulators is another key component of
[43:05] modulators is another key component of all of this and as I mentioned we're
[43:07] all of this and as I mentioned we're talking about the disk modulators. Yeah.
[43:09] talking about the disk modulators. Yeah. Yeah.
[43:13] >> The power goes up. Yeah. the number of
[43:15] >> The power goes up. Yeah. the number of channels uh goes down because
[43:18] channels uh goes down because >> the Q factor
[43:19] >> the Q factor >> the Q factor
[43:22] >> the Q factor the Q factor is a it it uh it's it's
[43:26] the Q factor is a it it uh it's it's just a matter of how many pulses you
[43:28] just a matter of how many pulses you have running around in in in that in
[43:29] have running around in in in that in that comb generator. So I don't think
[43:31] that comb generator. So I don't think the the device stays the same. Um but
[43:34] the the device stays the same. Um but the different FSR states and and so
[43:37] the different FSR states and and so having uh what does happen though which
[43:39] having uh what does happen though which is I don't exactly know why but as you
[43:43] is I don't exactly know why but as you go to a larger FSR the the conversion
[43:47] go to a larger FSR the the conversion efficiency from the pump to the comb
[43:49] efficiency from the pump to the comb goes up. It goes up which is very good
[43:52] goes up. It goes up which is very good because the higher the the conversion
[43:54] because the higher the the conversion efficiency the more energy efficient the
[43:56] efficiency the more energy efficient the link is. That's very very important
[43:58] link is. That's very very important part. So operating at the 300 gigahertz
[44:01] part. So operating at the 300 gigahertz is much more efficient. Uh we have a
[44:03] is much more efficient. Uh we have a recent we have an upcoming cleo paper on
[44:05] recent we have an upcoming cleo paper on that where we show that you get much
[44:07] that where we show that you get much better links with with the larger FSR.
[44:10] better links with with the larger FSR. Yeah, just a little little note. The
[44:12] Yeah, just a little little note. The discs the these are the disk modulators.
[44:15] discs the these are the disk modulators. We went through multiple generations of
[44:17] We went through multiple generations of of these disc. So what what's important
[44:19] of these disc. So what what's important about these discs? There are vertical
[44:21] about these discs? There are vertical dope junctions. Um and so the the
[44:25] dope junctions. Um and so the the challenge there is you want to design it
[44:27] challenge there is you want to design it so that it's single mode, right? So
[44:29] so that it's single mode, right? So obviously the the whispering gallery is
[44:32] obviously the the whispering gallery is not a single mode. It's the disk is not
[44:34] not a single mode. It's the disk is not a single mode device. So it's a
[44:36] a single mode device. So it's a challenge. You want to design it so that
[44:38] challenge. You want to design it so that everything has loss as much as possible
[44:41] everything has loss as much as possible except the fundamental mode. And then
[44:44] except the fundamental mode. And then you want to design it to go as fast as
[44:46] you want to design it to go as fast as you can with as low voltage as you can.
[44:49] you can with as low voltage as you can. Oh, and then you need to have thermal
[44:51] Oh, and then you need to have thermal control all in the smallest possible
[44:53] control all in the smallest possible disc that you can. Um, so our are our so
[44:57] disc that you can. Um, so our are our so this is kind of where we're driving and
[44:59] this is kind of where we're driving and we're able to get kind of record small
[45:01] we're able to get kind of record small discs. Um, this one this FSR
[45:05] discs. Um, this one this FSR so you want to get the largest FSR that
[45:07] so you want to get the largest FSR that you can on the disc and the best that
[45:10] you can on the disc and the best that we've gotten so far is about 60
[45:13] we've gotten so far is about 60 nanometers which is very nice. We can
[45:15] nanometers which is very nice. We can cover almost the entire comb in that.
[45:17] cover almost the entire comb in that. And then we have new designs where we
[45:21] And then we have new designs where we call them FSR less FSR less combs that
[45:25] call them FSR less FSR less combs that just have a single resonance and by
[45:27] just have a single resonance and by designing
[45:29] designing uh by I I don't have it in the stock but
[45:31] uh by I I don't have it in the stock but basically by designing the coupler this
[45:33] basically by designing the coupler this coupler over here the curve coupler you
[45:36] coupler over here the curve coupler you can design it so that through dispersion
[45:38] can design it so that through dispersion engineering so that only one of the FSR
[45:41] engineering so that only one of the FSR lines survives. So we're still we're
[45:44] lines survives. So we're still we're still on that. But so this is but I want
[45:47] still on that. But so this is but I want to give you understand where we're
[45:48] to give you understand where we're going. And the smallest that we can make
[45:51] going. And the smallest that we can make it is is about uh two two microns uh two
[45:55] it is is about uh two two microns uh two microns radius. So it's it's small but
[45:58] microns radius. So it's it's small but we hope to get smaller. Obviously
[46:00] we hope to get smaller. Obviously there's a limit of the curvature. Maybe
[46:02] there's a limit of the curvature. Maybe we can get a solution for for the losses
[46:04] we can get a solution for for the losses there. Um but here's here's another
[46:07] there. Um but here's here's another important thing. We can get modulation
[46:09] important thing. We can get modulation pretty fast. 32 gig um with only 04
[46:14] pretty fast. 32 gig um with only 04 volts peak-to peak. So that's important
[46:16] volts peak-to peak. So that's important because it's really CMOS compatible
[46:19] because it's really CMOS compatible advanced node compatible. Um sometimes
[46:22] advanced node compatible. Um sometimes people will say, "Oh, it's made up in
[46:24] people will say, "Oh, it's made up in it's silicon photonics, so it's CMOS
[46:26] it's silicon photonics, so it's CMOS compatible." Okay, but it's not really
[46:28] compatible." Okay, but it's not really CMOS compatible if you need one and a
[46:30] CMOS compatible if you need one and a half volts or three volts to drive it.
[46:33] half volts or three volts to drive it. Okay, that's not advanced CMOS
[46:34] Okay, that's not advanced CMOS compatible. Advanced CMOS has very low
[46:37] compatible. Advanced CMOS has very low voltages that you need to work with.
[46:39] voltages that you need to work with. Um, again, we're we're like really
[46:42] Um, again, we're we're like really focused on on the energy, right? Another
[46:46] focused on on the energy, right? Another big component of the energy is is the
[46:48] big component of the energy is is the thermal control that you need to apply
[46:50] thermal control that you need to apply to everything. And so we worked with AIM
[46:53] to everything. And so we worked with AIM on uh introducing undercut. So basically
[46:57] on uh introducing undercut. So basically drilling holes in the wafer, right? So
[46:59] drilling holes in the wafer, right? So you have air air pockets underneath your
[47:02] you have air air pockets underneath your devices. So this is an example. So
[47:05] devices. So this is an example. So without the undercut
[47:07] without the undercut um here's you apply zero one mill 1 mill
[47:11] um here's you apply zero one mill 1 mill watt 2 millwatts three four five and you
[47:13] watt 2 millwatts three four five and you tune you tune the discs a little bit
[47:16] tune you tune the discs a little bit with the undercut you apply the same
[47:19] with the undercut you apply the same amount of power you can tune a lot more
[47:22] amount of power you can tune a lot more right so much much more energy efficient
[47:25] right so much much more energy efficient uh and this is about 5x because the disc
[47:28] uh and this is about 5x because the disc is so small with larger devices we can
[47:31] is so small with larger devices we can get up to 20x and and higher so this is
[47:34] get up to 20x and and higher so this is another another important uh part.
[47:38] another another important uh part. Uh lastly with the discs um this is what
[47:42] Uh lastly with the discs um this is what we learned from being able to do the
[47:43] we learned from being able to do the wafer scale fabrication right so when
[47:46] wafer scale fabrication right so when you do wafer scale you look at all the
[47:47] you do wafer scale you look at all the variations and if you design rings
[47:52] variations and if you design rings uh you end up having the rings have a
[47:54] uh you end up having the rings have a lot more variation in their resonant
[47:57] lot more variation in their resonant frequency versus discs and of course
[47:59] frequency versus discs and of course it's very simple right the discs only
[48:01] it's very simple right the discs only have a one one diameter that you have to
[48:04] have a one one diameter that you have to that you care about right that you
[48:06] that you care about right that you design for a specific frequency
[48:08] design for a specific frequency With the rings, you have to design the
[48:10] With the rings, you have to design the wave guide. You have two edges that you
[48:12] wave guide. You have two edges that you care about. So, at the end of the day,
[48:15] care about. So, at the end of the day, you design a bunch of rings for your
[48:16] you design a bunch of rings for your link and then you make the measurements
[48:19] link and then you make the measurements and the resonance are like on top of
[48:21] and the resonance are like on top of each other, not where you design them.
[48:23] each other, not where you design them. It's very it's very depressing and then
[48:25] It's very it's very depressing and then you have to tune them over. The discs
[48:29] you have to tune them over. The discs you design them,
[48:31] you design them, they're very close, right? So, this is
[48:33] they're very close, right? So, this is fresh. So this is in this example we
[48:36] fresh. So this is in this example we have uh 64 discs. It's actually more
[48:39] have uh 64 discs. It's actually more than 64 because we have an extra one. So
[48:41] than 64 because we have an extra one. So four time 17 instead of four* 16 in this
[48:45] four time 17 instead of four* 16 in this this particular link. And look this is
[48:48] this particular link. And look this is measured no thermal control just out.
[48:51] measured no thermal control just out. And so now we have to do just a little
[48:53] And so now we have to do just a little bit of thermal control to tune them
[48:55] bit of thermal control to tune them over. And that saves a lot of energy for
[48:57] over. And that saves a lot of energy for us as well.
[48:59] us as well. All right. Um uh so we we we want to go
[49:04] All right. Um uh so we we we want to go to the 3D integration because we want to
[49:06] to the 3D integration because we want to combine this photonic chip with the
[49:08] combine this photonic chip with the electronics
[49:10] electronics uh to be able to get the high bandwidth
[49:12] uh to be able to get the high bandwidth density and so forth. Um and so this is
[49:15] density and so forth. Um and so this is um this is work that uh we did uh also
[49:18] um this is work that uh we did uh also in collaboration with um Cornell
[49:20] in collaboration with um Cornell University. There's a uh Alan uh Alan
[49:24] University. There's a uh Alan uh Alan Molnar is a professor there. He's a
[49:26] Molnar is a professor there. He's a electronic designer. So he designed a
[49:28] electronic designer. So he designed a simple EIC for us uh in 28 nanometer
[49:32] simple EIC for us uh in 28 nanometer TSMC very very straightforward node and
[49:35] TSMC very very straightforward node and then we did this this package where it's
[49:38] then we did this this package where it's basically the EIC on top of the on top
[49:41] basically the EIC on top of the on top of the pick. So this was the first real
[49:44] of the pick. So this was the first real demonstration of 3D. Yeah.
[49:47] demonstration of 3D. Yeah. And means
[49:48] And means >> of course I'm so sorry. I'm so sorry. Um
[49:51] >> of course I'm so sorry. I'm so sorry. Um so EIC is the electronic chip electronic
[49:55] so EIC is the electronic chip electronic integrated circuit. Uh the pick is the
[49:57] integrated circuit. Uh the pick is the fatonic uh fatonic integrated circuit.
[50:00] fatonic uh fatonic integrated circuit. I'm I'm I'm sorry for that. Um and so
[50:02] I'm I'm I'm sorry for that. Um and so basically you you design this this
[50:05] basically you you design this this particular electronic chip. You send it
[50:09] particular electronic chip. You send it to TSMC to be fabricated. It comes back.
[50:12] to TSMC to be fabricated. It comes back. You have to do there's a lot of work
[50:14] You have to do there's a lot of work that you need to do here at the
[50:16] that you need to do here at the interface so that they can they can bond
[50:18] interface so that they can they can bond to each other. Um and the um these uh uh
[50:24] to each other. Um and the um these uh uh and and we're trying to make everything
[50:26] and and we're trying to make everything as dense as we possibly ca can. So the
[50:28] as dense as we possibly ca can. So the pads are about 25 microns apart from
[50:32] pads are about 25 microns apart from each other. So this has to be done with
[50:34] each other. So this has to be done with pretty high precision. Um and so this
[50:37] pretty high precision. Um and so this was this was a like a really exciting
[50:40] was this was a like a really exciting attempt to really do this uh together.
[50:43] attempt to really do this uh together. And um uh so so this this this work this
[50:48] And um uh so so this this this work this is actually the chip that this this is
[50:50] is actually the chip that this this is actually the electronic chip on top of
[50:53] actually the electronic chip on top of the phatonic chip 3D integrated with all
[50:56] the phatonic chip 3D integrated with all the devices. In this case we have 80
[50:59] the devices. In this case we have 80 channels 80 channels. So this is this is
[51:03] channels 80 channels. So this is this is the chip right here. Uh this is uh a
[51:06] the chip right here. Uh this is uh a quarter uh US quarter. Um and uh look
[51:11] quarter uh US quarter. Um and uh look look at this precision of of the
[51:13] look at this precision of of the assembly. Right? So these are 25 microns
[51:16] assembly. Right? So these are 25 microns apart. Um before we succeeded in doing
[51:20] apart. Um before we succeeded in doing this, we tried to do it with not 300
[51:23] this, we tried to do it with not 300 millimeter wafers and we couldn't get it
[51:25] millimeter wafers and we couldn't get it to work because like there's no way they
[51:27] to work because like there's no way they can get the alignment to work. Um you
[51:30] can get the alignment to work. Um you can do it even better. You know 10 10
[51:33] can do it even better. You know 10 10 microns is pretty good. um these days
[51:36] microns is pretty good. um these days they can they can do better with hybrid
[51:38] they can they can do better with hybrid bonding and things like that. So there's
[51:40] bonding and things like that. So there's even you can get even even denser but
[51:42] even you can get even even denser but this was great and this in this chip um
[51:46] this was great and this in this chip um we were able to get uh just over five
[51:49] we were able to get uh just over five terabit uh per per millimeter uh in this
[51:52] terabit uh per per millimeter uh in this case per millimeter squared but five
[51:54] case per millimeter squared but five terabit uh with with this very high
[51:57] terabit uh with with this very high density uh uh chip. Um so we took a
[52:00] density uh uh chip. Um so we took a picture of it. I I was giving a talk and
[52:02] picture of it. I I was giving a talk and we just got the results and so I I
[52:05] we just got the results and so I I wanted to get a picture of it from the
[52:07] wanted to get a picture of it from the students and so the it was a young
[52:09] students and so the it was a young student early junior student uh in the
[52:13] student early junior student uh in the lab and he said he said sure I'll take a
[52:15] lab and he said he said sure I'll take a picture and he put this quarter right
[52:18] picture and he put this quarter right here in for the picture right so that
[52:20] here in for the picture right so that you get a sense of how big it is and I
[52:23] you get a sense of how big it is and I almost had a heart attack right because
[52:25] almost had a heart attack right because he put the quarter right on top of the
[52:26] he put the quarter right on top of the wire bonds
[52:28] wire bonds so but everything was Okay. Everything
[52:32] so but everything was Okay. Everything was okay. It worked. It was okay. It was
[52:34] was okay. It worked. It was okay. It was okay. Uh he survived. Um so this was
[52:38] okay. Uh he survived. Um so this was great.
[52:40] great. >> So student survived. Yeah. Yeah. The
[52:42] >> So student survived. Yeah. Yeah. The student survived and he went on to to do
[52:44] student survived and he went on to to do good good things. Yeah. Yeah. Um so so
[52:48] good good things. Yeah. Yeah. Um so so this was great. So this got the
[52:49] this was great. So this got the bandwidth density, right? What about the
[52:51] bandwidth density, right? What about the energy? What about the energy? Right. We
[52:53] energy? What about the energy? Right. We keep talking about the energy. So this
[52:56] keep talking about the energy. So this is by the way this is all 80 channels.
[52:59] is by the way this is all 80 channels. In this case there are only 10 Gbit per
[53:01] In this case there are only 10 Gbit per second uh working but because the
[53:04] second uh working but because the circuit was only able to drive 10 Gbit
[53:06] circuit was only able to drive 10 Gbit per second now we can do 32 Gbits. But
[53:09] per second now we can do 32 Gbits. But look at that 80 channels at 10 Gbit per
[53:12] look at that 80 channels at 10 Gbit per second and only 50 phento jewels per bit
[53:16] second and only 50 phento jewels per bit for the entire the entire chip the
[53:18] for the entire the entire chip the entire transmitter part of the chip. And
[53:22] entire transmitter part of the chip. And uh so this was this was a this is a
[53:23] uh so this was this was a this is a record uh on on the TX side. We then
[53:28] record uh on on the TX side. We then looked at the RX side. So on the RX
[53:31] looked at the RX side. So on the RX side, you know, the the detectors and
[53:36] side, you know, the the detectors and basically the bit error rate tester is
[53:38] basically the bit error rate tester is on the chip. It's designed to be on on
[53:40] on the chip. It's designed to be on on the electronic side of the chip. So it's
[53:42] the electronic side of the chip. So it's it's it's getting the data right there,
[53:44] it's it's getting the data right there, but it has a limited amount of memory
[53:47] but it has a limited amount of memory because of the simple simplicity of the
[53:49] because of the simple simplicity of the chip. So we were only able to measure 10
[53:51] chip. So we were only able to measure 10 the minus 10 instead of 10 the minus12
[53:53] the minus 10 instead of 10 the minus12 that's okay but look look at these eyes
[53:56] that's okay but look look at these eyes just you could drive a bus through these
[53:58] just you could drive a bus through these eyes and so this is just under seven 70
[54:04] eyes and so this is just under seven 70 phantomjles per bit this is the entire
[54:06] phantomjles per bit this is the entire receiver the TIA the front end
[54:08] receiver the TIA the front end everything so TX
[54:12] everything so TX at five terabit per millimeter squared
[54:15] at five terabit per millimeter squared 120 ftojles per bit um so we were very
[54:19] 120 ftojles per bit um so we were very excited about this result. You know,
[54:20] excited about this result. You know, we're really uh going there. Um I see
[54:24] we're really uh going there. Um I see that we're getting close to the end of
[54:25] that we're getting close to the end of the hour, so I'll speed up a little bit.
[54:27] the hour, so I'll speed up a little bit. Um so now, you know, we want to do more
[54:31] Um so now, you know, we want to do more channels. You know, we care. One of the
[54:33] channels. You know, we care. One of the other things we care about is is the
[54:35] other things we care about is is the losses. Everything. Anything that you
[54:37] losses. Everything. Anything that you can do to reduce the losses is a win.
[54:40] can do to reduce the losses is a win. Anything. Uh we just we're fighting
[54:42] Anything. Uh we just we're fighting losses because the less loss, the better
[54:45] losses because the less loss, the better the energy consumption. Um so we do
[54:48] the energy consumption. Um so we do that. This is this is more recent result
[54:52] that. This is this is more recent result with the higher speeds. So the higher
[54:54] with the higher speeds. So the higher speeds are going to consume more energy
[54:56] speeds are going to consume more energy because the electronics just takes more
[54:58] because the electronics just takes more energy. So it's a little bit different
[55:01] energy. So it's a little bit different um uh uh tradeoff but we're still able
[55:06] um uh uh tradeoff but we're still able to get uh and this is including the
[55:08] to get uh and this is including the laser. So the laser also takes a certain
[55:10] laser. So the laser also takes a certain amount of energy because you have a wall
[55:12] amount of energy because you have a wall plug efficiency, you have the conversion
[55:14] plug efficiency, you have the conversion efficiency, everything needs to be taken
[55:15] efficiency, everything needs to be taken into account. Uh but we can we for the
[55:18] into account. Uh but we can we for the entire link we can get about 300
[55:20] entire link we can get about 300 fptojles per bit but this link is going
[55:23] fptojles per bit but this link is going is going uh much faster 32 gigabits per
[55:27] is going uh much faster 32 gigabits per um uh per second per channel. So this is
[55:31] um uh per second per channel. So this is uh more and more of the latest. This is
[55:34] uh more and more of the latest. This is an example of okay now what do we do
[55:36] an example of okay now what do we do right so here here is our chip here is
[55:39] right so here here is our chip here is the electronic chip the EI the pig and
[55:42] the electronic chip the EI the pig and the EIC now that I explained what they
[55:44] the EIC now that I explained what they are and now what we wanted to do was
[55:46] are and now what we wanted to do was show that you can actually so we we we
[55:48] show that you can actually so we we we we did so far we did just dummy data but
[55:51] we did so far we did just dummy data but now we wanted to show like real data can
[55:53] now we wanted to show like real data can you put like real compute data on it and
[55:56] you put like real compute data on it and so you know we don't have access to
[55:58] so you know we don't have access to integrating with the GPUs but we were
[56:01] integrating with the GPUs but we were able to work with Intel
[56:03] able to work with Intel to integrate with FPGA, right? FPGA. So,
[56:07] to integrate with FPGA, right? FPGA. So, this is an FPGA package um that's uh
[56:11] this is an FPGA package um that's uh where the entire package is is going is
[56:14] where the entire package is is going is it's almost done. We have half of it
[56:16] it's almost done. We have half of it done. Is going to be almost a 100
[56:18] done. Is going to be almost a 100 terabyte package.
[56:20] terabyte package. Um now the FPGA never in its life is
[56:23] Um now the FPGA never in its life is going to generate a 100 terabyte of
[56:25] going to generate a 100 terabyte of data. It's okay. But we just want to
[56:27] data. It's okay. But we just want to show that we can program the FPGA for
[56:30] show that we can program the FPGA for something and it comes out the optical
[56:32] something and it comes out the optical side. So that's the goal here. Um, okay.
[56:36] side. So that's the goal here. Um, okay. I'll I'll I'll skip this part, but this
[56:38] I'll I'll I'll skip this part, but this is about this is a little bit more at
[56:39] is about this is a little bit more at the back to the system level. So one of
[56:42] the back to the system level. So one of the challenges that we have that's
[56:45] the challenges that we have that's growing, you see this is from very
[56:46] growing, you see this is from very recent from like 20 2019 to today and
[56:50] recent from like 20 2019 to today and continuing to grow is the gap between
[56:53] continuing to grow is the gap between the compute and the memory. That's, you
[56:56] the compute and the memory. That's, you know, you just cannot put enough memory
[56:59] know, you just cannot put enough memory on this GPUs. There's not enough space.
[57:02] on this GPUs. There's not enough space. And everybody's fighting this.
[57:04] And everybody's fighting this. Everybody's fighting this. Um, and so,
[57:07] Everybody's fighting this. Um, and so, you know, you're trying to get uh so so
[57:10] you know, you're trying to get uh so so the thinking is what if we took the
[57:12] the thinking is what if we took the phatonics
[57:14] phatonics that we have, right? This is the current
[57:16] that we have, right? This is the current architecture. They put the memory all
[57:18] architecture. They put the memory all around the GPU, right? As much as they
[57:21] around the GPU, right? As much as they can.
[57:22] can. But what if we took that and we changed
[57:25] But what if we took that and we changed it? So
[57:27] it? So all around the GPU is just optics and
[57:31] all around the GPU is just optics and then you can go you're not limited by
[57:34] then you can go you're not limited by the package anymore. You could just go a
[57:35] the package anymore. You could just go a little further out and put as much
[57:37] little further out and put as much memory as you want. You have as much
[57:39] memory as you want. You have as much space as you want because you're not
[57:41] space as you want because you're not limited by the interposer by the
[57:43] limited by the interposer by the package. And so this is this is where
[57:45] package. And so this is this is where we're going now. Really in increasing
[57:48] we're going now. Really in increasing the best of both worlds, right? all the
[57:50] the best of both worlds, right? all the all the memory bandwidth that you want
[57:52] all the memory bandwidth that you want and the capacity. That's the key. Today
[57:56] and the capacity. That's the key. Today we're limited by the capacity. Um so
[57:59] we're limited by the capacity. Um so this is kind of the the system
[58:00] this is kind of the the system architecture. Um and these are these are
[58:04] architecture. Um and these are these are some initial results and we're we're
[58:06] some initial results and we're we're seeing that we can get you know
[58:09] seeing that we can get you know tremendous speed ups like maybe a factor
[58:11] tremendous speed ups like maybe a factor of three um maybe even even more than
[58:14] of three um maybe even even more than that speed up on on the workload. And
[58:17] that speed up on on the workload. And what this means is if you get a factor
[58:19] what this means is if you get a factor of three speed up, you've reduced your
[58:22] of three speed up, you've reduced your the energy that it takes to to to do
[58:25] the energy that it takes to to to do that computation by that factor, right?
[58:28] that computation by that factor, right? It's about, you know, how much training
[58:29] It's about, you know, how much training you can do per unit jewel. Okay, so let
[58:32] you can do per unit jewel. Okay, so let let's keep going. I'll I'll skip this
[58:34] let's keep going. I'll I'll skip this part. This is about optical switching.
[58:37] part. This is about optical switching. Uh but I'm happy to answer questions
[58:39] Uh but I'm happy to answer questions about that. We we're looking at one of
[58:41] about that. We we're looking at one of the key questions in sort of optical
[58:44] the key questions in sort of optical connected computing systems is you you
[58:47] connected computing systems is you you might be familiar with papers in the
[58:50] might be familiar with papers in the last years about we put optical switches
[58:53] last years about we put optical switches MEMS like Google has MEM switches in
[58:55] MEMS like Google has MEM switches in their in their systems right and nobody
[58:59] their in their systems right and nobody really knows like well how big do those
[59:02] really knows like well how big do those switches need to be how fast do they
[59:04] switches need to be how fast do they need to go Google Google is just using
[59:06] need to go Google Google is just using them to to reconfigure their system and
[59:08] them to to reconfigure their system and for reliability. But if you put the
[59:11] for reliability. But if you put the switches really inside the system like a
[59:13] switches really inside the system like a fabric, um it's not really clear. Do
[59:16] fabric, um it's not really clear. Do they need to be a nancond? That's bad
[59:18] they need to be a nancond? That's bad because if they need to be a nancond,
[59:20] because if they need to be a nancond, it's very hard to make. Imagine making a
[59:23] it's very hard to make. Imagine making a thousand by,000 optical switch with
[59:25] thousand by,000 optical switch with nancond reconfiguration.
[59:28] nancond reconfiguration. Very hard. And it's going to take a lot
[59:30] Very hard. And it's going to take a lot of energy too, which we can't have. You
[59:32] of energy too, which we can't have. You can make a thousand by,000 me switch.
[59:35] can make a thousand by,000 me switch. Yeah. But it can be like a millisecond.
[59:38] Yeah. But it can be like a millisecond. So, so this is this this particular
[59:40] So, so this is this this particular paper is examines that and has a great
[59:43] paper is examines that and has a great great answer on it. I'm happy to answer
[59:45] great answer on it. I'm happy to answer questions and so let me just uh finish
[59:47] questions and so let me just uh finish up. This is this is my world that we're
[59:51] up. This is this is my world that we're working in and and this is these are the
[59:54] working in and and this is these are the great uh people and I this is the new
[59:56] great uh people and I this is the new member of our group right here. His name
[59:59] member of our group right here. His name is Levon. So, thank you so much for
[01:00:01] is Levon. So, thank you so much for listening and I'm happy to answer any
[01:00:03] listening and I'm happy to answer any questions.
[01:00:12] THANKS a lot for this really impressive
[01:00:13] THANKS a lot for this really impressive talk and also the introduction in the
[01:00:15] talk and also the introduction in the beginning that's was very very um
[01:00:17] beginning that's was very very um helpful. Are there questions?
[01:00:22] >> I have this big I give you the first
[01:00:24] >> I have this big I give you the first one.
[01:00:27] >> Thanks a lot.
[01:00:28] >> Thanks a lot. >> Uh I would like to go back to the page
[01:00:30] >> Uh I would like to go back to the page where you introduced embedded photonics
[01:00:33] where you introduced embedded photonics as the best world of both. Yeah.
[01:00:36] as the best world of both. Yeah. >> From uh let's say the US American GPUs
[01:00:39] >> From uh let's say the US American GPUs and the uh Chinese optical
[01:00:42] and the uh Chinese optical interconnects.
[01:00:42] interconnects. >> Sure. Yeah.
[01:00:44] >> Sure. Yeah. >> I mean you have introduced it in the
[01:00:45] >> I mean you have introduced it in the slides afterwards to us but there seems
[01:00:47] slides afterwards to us but there seems to be something missing because
[01:00:48] to be something missing because otherwise Nvidia would be selling this
[01:00:50] otherwise Nvidia would be selling this to us.
[01:00:51] to us. >> What is missing? It doesn't ex it does
[01:00:55] >> What is missing? It doesn't ex it does not exist yet in in in large scale. Like
[01:00:59] not exist yet in in in large scale. Like Nvidia, believe me, they're intensely.
[01:01:01] Nvidia, believe me, they're intensely. You might have seen just a few days ago,
[01:01:04] You might have seen just a few days ago, they invested $2 billion each in in in
[01:01:08] they invested $2 billion each in in in coherent and and momentum. Uh they
[01:01:11] coherent and and momentum. Uh they invested more in Ayara Labs, another
[01:01:14] invested more in Ayara Labs, another startup company. Uh because
[01:01:17] startup company. Uh because to to make it to make it at large scale,
[01:01:20] to to make it to make it at large scale, the ecosystem is not there. Um, of
[01:01:23] the ecosystem is not there. Um, of course there's some technical challenges
[01:01:24] course there's some technical challenges that still need to be solved uh for
[01:01:26] that still need to be solved uh for real, but if if it was there and it was
[01:01:29] real, but if if it was there and it was at at the cost point where they they
[01:01:32] at at the cost point where they they they need to have it, they would do it.
[01:01:34] they need to have it, they would do it. And they're they're definitely marching
[01:01:36] And they're they're definitely marching in that direction. The other thing
[01:01:37] in that direction. The other thing that's also important to keep in mind is
[01:01:39] that's also important to keep in mind is that, you know, they're the leader and
[01:01:42] that, you know, they're the leader and so they get to decide what the what the
[01:01:46] so they get to decide what the what the pace is. You know, do they need to spend
[01:01:49] pace is. You know, do they need to spend a lot of money on an expensive optical
[01:01:51] a lot of money on an expensive optical system? Maybe not yet until there's
[01:01:54] system? Maybe not yet until there's competition. So I think that's that's
[01:01:56] competition. So I think that's that's kind of where things are.
[01:01:58] kind of where things are. >> Yeah, absolutely. But we're gonna make
[01:02:00] >> Yeah, absolutely. But we're gonna make it.
[01:02:01] it. >> Yeah.
[01:02:02] >> Yeah. >> All right. Thank you for your talk. So
[01:02:04] >> All right. Thank you for your talk. So my question is at the end of your talk
[01:02:06] my question is at the end of your talk you you uh suggested that we could just
[01:02:09] you you uh suggested that we could just uh replace uh those HPM stacks
[01:02:11] uh replace uh those HPM stacks surrounding your GPU with optical
[01:02:14] surrounding your GPU with optical interconnects. But I think one of the
[01:02:15] interconnects. But I think one of the reasons why those HPM stacks are so
[01:02:17] reasons why those HPM stacks are so close is is just latency.
[01:02:19] close is is just latency. >> Latency is important. Can you comment on
[01:02:21] >> Latency is important. Can you comment on latency with your optics techniques?
[01:02:24] latency with your optics techniques? >> Yes. Yes. Absolutely. Latency is
[01:02:25] >> Yes. Yes. Absolutely. Latency is important for sure. Uh lat it also
[01:02:28] important for sure. Uh lat it also depends a little bit on on on the
[01:02:31] depends a little bit on on on the application. So first of all, these are
[01:02:33] application. So first of all, these are not going to be that far away, right?
[01:02:35] not going to be that far away, right? The you know, right now, you know,
[01:02:38] The you know, right now, you know, you're always going to have some local
[01:02:40] you're always going to have some local memory. You have to have that for sure.
[01:02:42] memory. You have to have that for sure. Um but you know, the you know, if you're
[01:02:45] Um but you know, the you know, if you're going to go, you know, a few more
[01:02:46] going to go, you know, a few more cycles, you know, off the chip, right?
[01:02:49] cycles, you know, off the chip, right? The idea is that we can um you know
[01:02:52] The idea is that we can um you know these guys are really limited by by how
[01:02:55] these guys are really limited by by how far you can you can have these
[01:02:57] far you can you can have these electrical um you know traces go. But if
[01:03:01] electrical um you know traces go. But if we if we go here
[01:03:03] we if we go here um you know can go any distance. Let's
[01:03:07] um you know can go any distance. Let's say let's say we just want to keep the
[01:03:09] say let's say we just want to keep the distances similar chip to chip. Now we
[01:03:12] distances similar chip to chip. Now we can this can be much much more bandwidth
[01:03:15] can this can be much much more bandwidth density and we can we can now stack HBMs
[01:03:19] density and we can we can now stack HBMs on another on a whole another uh uh
[01:03:22] on another on a whole another uh uh package and and have much more access to
[01:03:25] package and and have much more access to more. Yes, you're right. There's there's
[01:03:27] more. Yes, you're right. There's there's time offlight latency that that's going
[01:03:28] time offlight latency that that's going to be taken into account and the the
[01:03:31] to be taken into account and the the work that I went through very quickly,
[01:03:33] work that I went through very quickly, you know, takes that into account and
[01:03:35] you know, takes that into account and that's why it's not perfect. uh but you
[01:03:37] that's why it's not perfect. uh but you still get a benefit and the good thing
[01:03:40] still get a benefit and the good thing about um so latency is now the new name
[01:03:42] about um so latency is now the new name of the game uh and we're very focused on
[01:03:45] of the game uh and we're very focused on that um uh even even more than we were
[01:03:48] that um uh even even more than we were before first it was bandwidth you know
[01:03:50] before first it was bandwidth you know we got the bandwidth I think we can we
[01:03:52] we got the bandwidth I think we can we can get there and now let's let's
[01:03:53] can get there and now let's let's squeeze out the latency yeah
[01:03:57] >> thank you very much for this talk oh
[01:03:59] >> thank you very much for this talk oh okay uh I just wanted to know uh you
[01:04:02] okay uh I just wanted to know uh you talked about modulation on each channel
[01:04:04] talked about modulation on each channel right I wanted to know more how Did you
[01:04:06] right I wanted to know more how Did you get this thermal tuning?
[01:04:08] get this thermal tuning? >> How did I get the Sorry, the what
[01:04:10] >> How did I get the Sorry, the what tuning?
[01:04:10] tuning? >> Uh the thermal tuning. I thought like
[01:04:11] >> Uh the thermal tuning. I thought like you used thermal tuning.
[01:04:12] you used thermal tuning. >> Oh, thermal tuning. Yes. Yes. So, uh
[01:04:15] >> Oh, thermal tuning. Yes. Yes. So, uh each each one of those uh devices has
[01:04:20] each each one of those uh devices has uh
[01:04:22] uh Yeah. So, so they have um in it depends
[01:04:26] Yeah. So, so they have um in it depends on the process technology. Sometimes you
[01:04:28] on the process technology. Sometimes you can have metal heaters. In our case, in
[01:04:31] can have metal heaters. In our case, in our technology, we have doped doped
[01:04:33] our technology, we have doped doped heaters that we designed. So here's
[01:04:35] heaters that we designed. So here's here's an example of a filter. That's
[01:04:37] here's an example of a filter. That's the easy one to make. And so this is
[01:04:39] the easy one to make. And so this is this is the disc and it has this is a
[01:04:42] this is the disc and it has this is a doped heater that's sitting here and and
[01:04:45] doped heater that's sitting here and and then uh you know it has some contacts
[01:04:47] then uh you know it has some contacts and basically that's that's how the it's
[01:04:49] and basically that's that's how the it's a very very small amount of current that
[01:04:51] a very very small amount of current that we apply to it and that tunes the the
[01:04:54] we apply to it and that tunes the the device. It's more complicated when you
[01:04:57] device. It's more complicated when you have the modulator because you have both
[01:05:00] have the modulator because you have both RF that you're trying to get as fast as
[01:05:02] RF that you're trying to get as fast as you can and so you want to have as many
[01:05:03] you can and so you want to have as many contacts to minimize the the losses and
[01:05:07] contacts to minimize the the losses and somewhere you want to still get the
[01:05:08] somewhere you want to still get the heater. And so what we've done I didn't
[01:05:11] heater. And so what we've done I didn't go into it but we developed a half etch
[01:05:14] go into it but we developed a half etch technology where we can put the heater
[01:05:17] technology where we can put the heater outside the disc. So now we can make the
[01:05:19] outside the disc. So now we can make the disc a lot smaller. Uh, and still, you
[01:05:22] disc a lot smaller. Uh, and still, you know, we still need room for the RF, but
[01:05:24] know, we still need room for the RF, but we can make it a lot smaller and and the
[01:05:26] we can make it a lot smaller and and the the heater is outside the disc, uh, with
[01:05:29] the heater is outside the disc, uh, with a half edge, so it doesn't it doesn't
[01:05:31] a half edge, so it doesn't it doesn't overlap with the with the RF signals.
[01:05:33] overlap with the with the RF signals. Yeah,
[01:05:34] Yeah, it's it's like a two and a half, I would
[01:05:37] it's it's like a two and a half, I would say. Not exactly 3D. Yeah, like two and
[01:05:39] say. Not exactly 3D. Yeah, like two and a half. Yeah. Yeah, exactly. But
[01:05:42] a half. Yeah. Yeah, exactly. But definitely some design design
[01:05:45] definitely some design design innovations and process innovations.
[01:05:47] innovations and process innovations. This is why I was saying it's so
[01:05:48] This is why I was saying it's so important to I mean we're very fortunate
[01:05:50] important to I mean we're very fortunate to have this relationship with a foundry
[01:05:53] to have this relationship with a foundry where we can say I mean it's not that
[01:05:55] where we can say I mean it's not that simple but we can say oh let's let's do
[01:05:57] simple but we can say oh let's let's do this and let's do that uh and they they
[01:06:00] this and let's do that uh and they they can they can sort of do it. Yeah.
[01:06:04] can they can sort of do it. Yeah. >> Um my question is related to the last
[01:06:06] >> Um my question is related to the last one.
[01:06:06] one. >> Yeah.
[01:06:07] >> Yeah. >> You showed this undercut.
[01:06:08] >> You showed this undercut. >> Yes. Have this larger shift. So the
[01:06:10] >> Yes. Have this larger shift. So the reason must be that because with the
[01:06:12] reason must be that because with the undercut you reduce the heat
[01:06:14] undercut you reduce the heat conductivity.
[01:06:14] conductivity. >> Exactly. Exactly. That should also slow
[01:06:17] >> Exactly. Exactly. That should also slow down.
[01:06:17] down. >> Yes.
[01:06:17] >> Yes. >> So there's a there's a trade-off, right?
[01:06:20] >> So there's a there's a trade-off, right? >> Yes. Yes. It I think it actually
[01:06:23] >> Yes. Yes. It I think it actually >> I thought it speeds it. No, it slows it
[01:06:25] >> I thought it speeds it. No, it slows it down. It slows it down, but it's it's
[01:06:27] down. It slows it down, but it's it's really not not a big factor because the
[01:06:30] really not not a big factor because the the thermal is slow anyway. It's fine.
[01:06:33] the thermal is slow anyway. It's fine. And as long as we the thermal drift, the
[01:06:36] And as long as we the thermal drift, the thermal variations that that you're
[01:06:38] thermal variations that that you're going to what what do we worry about? we
[01:06:40] going to what what do we worry about? we worry about, you know, the the
[01:06:42] worry about, you know, the the environment and and so the the GPUs are
[01:06:44] environment and and so the the GPUs are somewhere. I mean, the GPUs have such,
[01:06:48] somewhere. I mean, the GPUs have such, you know, they they produce so much so
[01:06:51] you know, they they produce so much so much heat and so much power. Um, which
[01:06:54] much heat and so much power. Um, which is fine, which is fine. And they're
[01:06:56] is fine, which is fine. And they're cooled, right? They're water cooled and
[01:06:58] cooled, right? They're water cooled and and liquid coil, all kinds of that. What
[01:07:00] and liquid coil, all kinds of that. What what we actually care about the
[01:07:02] what we actually care about the phatonics is not actually the the
[01:07:04] phatonics is not actually the the temperature itself, but how fast does it
[01:07:06] temperature itself, but how fast does it move, right?
[01:07:08] move, right? And it doesn't move that fast. It's not
[01:07:10] And it doesn't move that fast. It's not gonna it doesn't change like you know
[01:07:13] gonna it doesn't change like you know just because of the physics. It's not
[01:07:14] just because of the physics. It's not going to change on you know more than
[01:07:16] going to change on you know more than like sort of tens of microsconds at
[01:07:19] like sort of tens of microsconds at best. And so this this is not a big
[01:07:20] best. And so this this is not a big factor but you're right that's a
[01:07:22] factor but you're right that's a sacrifice. That's very good point. It
[01:07:23] sacrifice. That's very good point. It was a trade-off. Yeah.
[01:07:25] was a trade-off. Yeah. >> Let me take a question from from Zoom.
[01:07:28] >> Let me take a question from from Zoom. >> Oh good.
[01:07:28] >> Oh good. >> So um how are the lasers connected to
[01:07:30] >> So um how are the lasers connected to the photonic chip? Photonic wire
[01:07:32] the photonic chip? Photonic wire bonding.
[01:07:34] bonding. >> Wire bonding. How how are they connected
[01:07:36] >> Wire bonding. How how are they connected to the chip?
[01:07:37] to the chip? >> Oh, the wire buns I'm sorry. Um yeah, so
[01:07:39] >> Oh, the wire buns I'm sorry. Um yeah, so the wire bonds were used to just uh
[01:07:42] the wire bonds were used to just uh connect to the circuit board. Um so
[01:07:45] connect to the circuit board. Um so there we used so in in the chip that we
[01:07:48] there we used so in in the chip that we made um we we had the the the
[01:07:52] made um we we had the the the electronics is 3D integrated. So the
[01:07:54] electronics is 3D integrated. So the high-speed electronics is right on top
[01:07:56] high-speed electronics is right on top of the photonic. So it's direct um
[01:07:59] of the photonic. So it's direct um bonded, you know, the pads are bonded.
[01:08:01] bonded, you know, the pads are bonded. But then we also have some thermal
[01:08:03] But then we also have some thermal control that wasn't on the integrated
[01:08:06] control that wasn't on the integrated electronic chip just in a in a circuit
[01:08:08] electronic chip just in a in a circuit board that we connected with wire bonds
[01:08:10] board that we connected with wire bonds to that. So there all all the wire bonds
[01:08:13] to that. So there all all the wire bonds are just connected to relatively slow uh
[01:08:15] are just connected to relatively slow uh circuitry for the control.
[01:08:17] circuitry for the control. >> Yeah.
[01:08:18] >> Yeah. >> So the and the second question from the
[01:08:19] >> So the and the second question from the same person Arthur Lobo how do you deal
[01:08:22] same person Arthur Lobo how do you deal with the thermal cross talk of heaters
[01:08:24] with the thermal cross talk of heaters slide 18. It's it's always a it's I I
[01:08:26] slide 18. It's it's always a it's I I get that question a lot and we actually
[01:08:28] get that question a lot and we actually do not really we do not see a big
[01:08:31] do not really we do not see a big problem. Um you know we have those we
[01:08:33] problem. Um you know we have those we have the discs you know they're they're
[01:08:35] have the discs you know they're they're very close to each other and I showed
[01:08:37] very close to each other and I showed you the very high density. Um you know
[01:08:40] you the very high density. Um you know they're 10 microns away from each other.
[01:08:42] they're 10 microns away from each other. We we really don't see I mean it's very
[01:08:44] We we really don't see I mean it's very very localized uh heating and it's
[01:08:46] very localized uh heating and it's actually not much heating. I mean the
[01:08:48] actually not much heating. I mean the chip is very cool right? you think about
[01:08:50] chip is very cool right? you think about the whole chip is spending 120
[01:08:53] the whole chip is spending 120 phentojles per per bit. Um that's that's
[01:08:56] phentojles per per bit. Um that's that's very it's a very cool uh chip. Of
[01:09:00] very it's a very cool uh chip. Of course, it's going to be eventually it's
[01:09:01] course, it's going to be eventually it's going to be near a a massive GPU. So
[01:09:05] going to be near a a massive GPU. So that's a different story, but the
[01:09:06] that's a different story, but the photonics is actually very very uh low
[01:09:10] photonics is actually very very uh low low power. Yeah.
[01:09:13] low power. Yeah. >> Yes. Thank you very Thank you very much
[01:09:14] >> Yes. Thank you very Thank you very much for the wonderful talk. Um I have a
[01:09:16] for the wonderful talk. Um I have a question on the power budget in your it
[01:09:20] question on the power budget in your it looked like the com generation is one of
[01:09:22] looked like the com generation is one of the biggest trends.
[01:09:22] the biggest trends. >> It is. Yeah.
[01:09:23] >> It is. Yeah. >> And maybe two questions for that. Did
[01:09:26] >> And maybe two questions for that. Did you are you considering or are you
[01:09:28] you are you considering or are you already using dark solid on coms that
[01:09:30] already using dark solid on coms that would be way more power efficient?
[01:09:32] would be way more power efficient? >> Uh yes. I mean I I I don't know the
[01:09:35] >> Uh yes. I mean I I I don't know the terminology as well. I I know Guyetta
[01:09:38] terminology as well. I I know Guyetta doesn't love to talk about dark
[01:09:39] doesn't love to talk about dark solatons. calls. So the the combs are
[01:09:43] solatons. calls. So the the combs are quote unquote normal GVD combs, right?
[01:09:46] quote unquote normal GVD combs, right? Normal group velocity uh dispersion
[01:09:47] Normal group velocity uh dispersion combs. Um and so the good thing about
[01:09:50] combs. Um and so the good thing about them is that you can have very high
[01:09:53] them is that you can have very high conversion efficiency, right? You're not
[01:09:55] conversion efficiency, right? You're not limited like the solaton to very low,
[01:09:58] limited like the solaton to very low, you know, lower conversion efficiency
[01:10:00] you know, lower conversion efficiency and you're not limited in the power that
[01:10:02] and you're not limited in the power that you can have. Um and that's those two
[01:10:05] you can have. Um and that's those two things are very very important. Now the
[01:10:07] things are very very important. Now the other contributors to this large energy
[01:10:10] other contributors to this large energy see now we we reduce other things like
[01:10:13] see now we we reduce other things like you know it's always like this game
[01:10:14] you know it's always like this game right the other things pop up like wow
[01:10:16] right the other things pop up like wow we didn't we didn't know the comb was so
[01:10:18] we didn't we didn't know the comb was so expensive energy-wise
[01:10:20] expensive energy-wise um so it's the pump laser the pump laser
[01:10:23] um so it's the pump laser the pump laser is a problem because we're just using
[01:10:26] is a problem because we're just using commercial pump right like a like a DFB
[01:10:30] commercial pump right like a like a DFB pump laser and we have to count
[01:10:32] pump laser and we have to count everything counts every jewel counts we
[01:10:35] everything counts every jewel counts we have to count what is the wall plug
[01:10:37] have to count what is the wall plug conversion efficiency. You know, you
[01:10:39] conversion efficiency. You know, you plug it to the wall, you take the
[01:10:41] plug it to the wall, you take the electrical, you know, current and you
[01:10:43] electrical, you know, current and you convert it to light. You can get that
[01:10:46] convert it to light. You can get that better. If you can get semiconductor
[01:10:48] better. If you can get semiconductor lasers to be 50% oil plug efficient,
[01:10:52] lasers to be 50% oil plug efficient, this number is going to go down a lot.
[01:10:53] this number is going to go down a lot. This number is going to go down. Yeah.
[01:10:55] This number is going to go down. Yeah. >> And kind of my additional question would
[01:10:56] >> And kind of my additional question would be how much can you reduce the power of
[01:10:58] be how much can you reduce the power of the com because if you have more
[01:11:00] the com because if you have more efficient photo detectors or
[01:11:02] efficient photo detectors or >> Yes, everything contributes. Yeah,
[01:11:04] >> Yes, everything contributes. Yeah, everything contributes. It's the photo
[01:11:06] everything contributes. It's the photo detectors. It's the sensitivity of the
[01:11:09] detectors. It's the sensitivity of the electronics as well. So, it's about
[01:11:11] electronics as well. So, it's about reducing the noise on the electronic
[01:11:12] reducing the noise on the electronic receiver. That's another big huge
[01:11:14] receiver. That's another big huge factor. And that's why as we go to
[01:11:16] factor. And that's why as we go to higher data rates, it's it becomes
[01:11:18] higher data rates, it's it becomes worse. That's why we were able to get
[01:11:20] worse. That's why we were able to get such good results at 10 gigabit per
[01:11:22] such good results at 10 gigabit per second because the sensitivity of the of
[01:11:25] second because the sensitivity of the of the electronics was very very good,
[01:11:27] the electronics was very very good, right? Low noise, low frequency, low
[01:11:29] right? Low noise, low frequency, low noise always. Um high frequency, you
[01:11:32] noise always. Um high frequency, you know, a little bit more noise. and and
[01:11:35] know, a little bit more noise. and and so um the way but the good thing about
[01:11:38] so um the way but the good thing about optics is you can increase the bandwidth
[01:11:41] optics is you can increase the bandwidth with with wavelengths and so that's
[01:11:43] with with wavelengths and so that's really kind of the design space so def
[01:11:46] really kind of the design space so def definitely anything you can do about the
[01:11:48] definitely anything you can do about the losses anything you can do about the
[01:11:50] losses anything you can do about the laser efficiency
[01:11:52] laser efficiency you know even fundamentally right I mean
[01:11:54] you know even fundamentally right I mean I I don't think you know right now these
[01:11:56] I I don't think you know right now these lasers are like maybe at best 25% wool
[01:12:00] lasers are like maybe at best 25% wool plug efficiency at best because we're
[01:12:03] plug efficiency at best because we're using high power is better and that's
[01:12:05] using high power is better and that's fine. Um, low power DFB lasers are
[01:12:09] fine. Um, low power DFB lasers are actually more like 5% or 3% well plug
[01:12:11] actually more like 5% or 3% well plug efficient. So it those are some some of
[01:12:15] efficient. So it those are some some of the considerations that are that are
[01:12:16] the considerations that are that are key.
[01:12:18] key. >> Yeah.
[01:12:19] >> Yeah. >> So I will take one more question from
[01:12:21] >> So I will take one more question from online and then um also um can you
[01:12:25] online and then um also um can you explain the yield numbers for pigs and
[01:12:26] explain the yield numbers for pigs and also for 3D integrated chips?
[01:12:29] also for 3D integrated chips? >> Uh the the
[01:12:30] >> Uh the the >> yield numbers
[01:12:31] >> yield numbers >> the yield. Oh, the yield um it's
[01:12:35] >> the yield. Oh, the yield um it's it's uh it's it's very good. I mean,
[01:12:38] it's uh it's it's very good. I mean, basically, you know, obviously we're not
[01:12:40] basically, you know, obviously we're not a commercial um entity. We're just doing
[01:12:43] a commercial um entity. We're just doing research, but you know, when when we
[01:12:46] research, but you know, when when we fabricate a wafer, we get six, you know,
[01:12:48] fabricate a wafer, we get six, you know, it has 64, you know, um we get 64 copies
[01:12:52] it has 64, you know, um we get 64 copies of everything. And there is you know the
[01:12:56] of everything. And there is you know the the yield is is you know basically every
[01:12:59] the yield is is you know basically every every you know every every uh diet
[01:13:01] every you know every every uh diet works. Um so the functionality is is
[01:13:04] works. Um so the functionality is is very high you know probably 90% and and
[01:13:07] very high you know probably 90% and and higher. The key is the variation. What's
[01:13:10] higher. The key is the variation. What's the variation across the wafer? um you
[01:13:13] the variation across the wafer? um you know and there's thickness variations
[01:13:16] know and there's thickness variations and and things like that that every
[01:13:18] and and things like that that every single thing affects the not so much the
[01:13:21] single thing affects the not so much the quality the devices work but exactly
[01:13:24] quality the devices work but exactly where the resonance is exactly where the
[01:13:26] where the resonance is exactly where the you know the the couplers are are
[01:13:28] you know the the couplers are are designed to um so yeah I mean we
[01:13:32] designed to um so yeah I mean we certainly get great yield for the
[01:13:34] certainly get great yield for the research that we want to do and and I I
[01:13:36] research that we want to do and and I I I would just stipulate that the
[01:13:39] I would just stipulate that the companies the commercial people are
[01:13:41] companies the commercial people are getting good yield build from the
[01:13:43] getting good yield build from the silicon fatonics for the commercial
[01:13:45] silicon fatonics for the commercial applications because they're making
[01:13:46] applications because they're making they're making products out of it. So
[01:13:49] they're making products out of it. So yeah,
[01:13:49] yeah, >> thank you.
[01:13:50] >> thank you. >> Are there more questions?
[01:13:52] >> Are there more questions? >> Good. And let's thank Oh, no.
[01:13:54] >> Good. And let's thank Oh, no. >> One more. Yes.
[01:13:58] >> One more. Yes. >> The the what?
[01:14:00] >> The the what? >> Oh, this we made this up. I had I had
[01:14:04] >> Oh, this we made this up. I had I had the this this goes back a little while.
[01:14:06] the this this goes back a little while. Um but this one is Colombia. This is the
[01:14:08] Um but this one is Colombia. This is the Colombia. They It's like a like a hat.
[01:14:11] Colombia. They It's like a like a hat. Yeah, that's the Colombia. The other
[01:14:13] Yeah, that's the Colombia. The other one, LRL, Lightwave Research Laboratory.
[01:14:16] one, LRL, Lightwave Research Laboratory. >> LRL.
[01:14:17] >> LRL. >> LRL. Yeah. Yeah.
[01:14:20] >> LRL. Yeah. Yeah. >> Okay. So, let's thank our speaker. Thank
[01:14:22] >> Okay. So, let's thank our speaker. Thank you.