Full Transcript
https://www.youtube.com/watch?v=0m1IvO5Hj2Y
[00:03] Uh thanks everyone for attending.
[00:05] I'm uh Dan at semi analysis.
[00:07] I'm the director of research.
[00:09] Um I also lead our networking practice.
[00:11] Um so you know we obviously have put out a lot of papers and newsletter and we also run market intelligence models.
[00:18] Uh we've we launched our AI accelerate AI networking model a few months ago and so you know we've certainly done a lot of work on CPO.
[00:26] Um you know so everything scale out scale up as well.
[00:28] So um I'll just quickly uh introduce today's panelists.
[00:30] We've got Adicha from GU, Lee Hong from ASC.
[00:36] Um uh CK from MediaTek and Eric from R Deluxe.
[00:38] But um I'll get them to introduce themselves uh one by one and then I'll go and ask a few questions.
[00:44] So maybe right to left.
[00:45] >> Oh, okay. Hi everyone.
[00:48] Uh thanks for coming.
[00:51] Uh my name is CK.
[00:54] I'm running the uh uh marketing in the media tech for the ASIC business.
[00:57] Good to good to be here.
[00:59] >> Okay. Hi, I'm Eric.
[01:01] I'm the CEO and founder of Artidilux.
[01:04] We are a company
[01:03] founder of Artidilux.
[01:03] We are a company specialized in Germanium based photonics.
[01:06] specialized in Germanium based photonics including the Germanian photo detector.
[01:08] including the Germanian photo detector APD with really high sensitivity and operate high temperature.
[01:11] APD with really high sensitivity and operate high temperature and um this ouration has been published few years ago in nature.
[01:15] ouration has been published few years ago in nature.
[01:15] So it's public data.
[01:17] ago in nature.
[01:17] So it's public data.
[01:17] welcome to take a look and also we are also working on Germania EM electro absorption modulator working on the OET not the conventional CN bit and yeah so that's why we have been working on recently.
[01:20] welcome to take a look and also we are also working on Germania EM electro.
[01:22] also working on Germania EM electro absorption modulator working on the OET.
[01:25] absorption modulator working on the OET not the conventional CN bit and yeah so.
[01:27] not the conventional CN bit and yeah so that's why we have been working on.
[01:29] that's why we have been working on recently and we are company now 10 years old about more than 100 people with a dedic dedicated team working on semiconductor optics and also IC design.
[01:32] recently and we are company now 10 years old about more than 100 people with a.
[01:34] old about more than 100 people with a dedic dedicated team working on.
[01:36] dedic dedicated team working on semiconductor optics and also IC design.
[01:39] semiconductor optics and also IC design yeah and uh I know to the topic we are going to so based on our innovation we are actually working on not just the wavegu base narrow and fast but also for the wine slow which is we call oroptics and we can share more during the later yeah later discussion.
[01:41] yeah and uh I know to the topic we are going to so based on our innovation we are actually working on not just the.
[01:43] going to so based on our innovation we are actually working on not just the wavegu base narrow and fast but also for.
[01:45] are actually working on not just the wavegu base narrow and fast but also for the wine slow which is we call oroptics.
[01:47] wavegu base narrow and fast but also for the wine slow which is we call oroptics and we can share more during the later yeah later discussion.
[01:50] the wine slow which is we call oroptics and we can share more during the later.
[01:51] and we can share more during the later yeah later discussion.
[01:53] yeah later discussion.
[01:53] Hi my name is Aditya uh I'm responsible for marketing at GU.
[01:56] Hi my name is Aditya uh I'm responsible for marketing at GU.
[02:00] GU is a design service company headquartered in Taiwan.
[02:02] service company headquartered in Taiwan.
[02:02] Uh we usually work really closely with
[02:04] Uh we usually work really closely with TSMC on all advanced uh packaging and advanced technologies.
[02:10] DSMC got into CPO so so did GC's and uh that's why we're here.
[02:15] Uh I'm Lee Hong Ka.
[02:19] I'm from ASC located a host Austin.
[02:20] I'm in charge of technology development and um promotion.
[02:25] Thank you.
[02:26] Yeah. Great. Thanks everyone.
[02:29] So I think what we'll start off with is um you know where you see the major needs and wants of customers on CPO.
[02:34] Um and you know perhaps at this point we can talk a bit distinctly about scale out CPO and scale up CPO where the needs and wants are different.
[02:40] Um I think much of our talk will probably focus on scale up.
[02:43] Um but you know can you just tell us what are the pain points that you're trying to solve for your customers and you know therefore what is your vision of the world particularly for scale up.
[02:55] Um so why don't we just go sort of one by one we can go left to right.
[02:58] Oh yes that's a very uh good question.
[03:01] I think uh that is a lot of people ask the same questions.
[03:04] So uh I just bring up
[03:06] Same questions.
[03:06] So uh I just bring up the first we see the needs and the first we see the needs and especially on the power and uh so the.
[03:12] Especially on the power and uh so the power hunter for data center so how we resolve it so the CPU is comes on the board and another things you know people.
[03:19] Board and another things you know people talk about the data rate when you hit 1.6T 6T and uh so currently for example um the copper interconnect over the plaable and is hit the bottom neck so we.
[03:32] Plaable and is hit the bottom neck so we see the need of CPO and definitely we also see the performance benefit the legacy uh improvement so we see a we are the ASC so we sure of a lot of customers.
[03:48] So we see the customer have a common u driving force about what we just mentioned.
[03:55] And is that mostly on the the um the CPO for for scale out domain that you're talking about or you also you know are think or is that comment in the context of scale up or scale out?
[04:06] Yeah.
[04:06] Yeah definitely uh we see the first you.
[04:08] Yeah definitely uh we see the first you know scaling up a lot of people talk about the scaling up and also the scaling out.
[04:13] So scaling out is more concerns about the long distance requirement.
[04:20] Yeah.
[04:22] Great. Thanks.
[04:22] Right. So uh to us the vision of CPO is to basically break the bottleneck set by copper.
[04:31] So uh if you attended the presentations thank you.
[04:35] So if you attended these presentations you kind of see that uh the demand for AI is insatiable.
[04:43] You need to connect more and more GPUs together.
[04:45] I think state-of-the-art right now is 72 GPUs.
[04:47] You can probably go to 144.
[04:51] you connect more of these uh you know power you start hitting the power wall you start liquid cooling these and you can keep doing this within a rack until the power basically explodes right so at some point you're going to have to go let's call it out of a rack I I don't want to use the term scale out because it just sounds too far out but just going from
[05:09] sounds too far out but just going from one rack to the other is to me uh.
[05:12] one rack to the other is to me uh basically going out of a rack so the.
[05:14] basically going out of a rack so the idea behind CPO is you can connect at.
[05:18] idea behind CPO is you can connect at this point maybe up to 124 four uh such.
[05:21] this point maybe up to 124 four uh such CPUs, GPUs, whatever you have it and uh.
[05:24] CPUs, GPUs, whatever you have it and uh it solves the the reach problem, the.
[05:26] it solves the the reach problem, the latency problem, the power problem.
[05:29] And let me try and contextualize this with.
[05:31] let me try and contextualize this with an example.
[05:34] So one of our photonix partners basically said that u so.
[05:37] partners basically said that u so everything is determined by tokens per.
[05:39] everything is determined by tokens per second tokconomics right so at this.
[05:42] point in the world what you're seeing is.
[05:44] a lot of interaction with chat GPT right.
[05:46] so it's human machine interaction so.
[05:48] so it's human machine interaction so when they generate tokens at 40 or 50.
[05:51] when they generate tokens at 40 or 50 tokens per second it's good enough.
[05:53] tokens per second it's good enough because you know a normal human can't.
[05:55] because you know a normal human can't read faster than that so this is good.
[05:57] read faster than that so this is good enough but as you start going into.
[06:00] enough but as you start going into agentic AI where this is not a human.
[06:02] agentic AI where this is not a human machine anymore.
[06:04] It's machine to machine.
[06:06] That's where the number of tokens goes to 200, 300, 400 tokens per.
[06:09] tokens goes to 200, 300, 400 tokens per second.
[06:09] And this is what is driving more.
[06:12] second.
[06:12] And this is what is driving more GPUs to be connected together.
[06:14] GPUs to be connected together.
[06:14] And this is where CPO comes in.
[06:17] is where CPO comes in.
[06:18] Yeah, I I think you know, one thing you're going to see very clearly on
[06:20] you're going to see very clearly on Inference Max, what we just launched
[06:22] Inference Max, what we just launched last week is you'll see a dramatic
[06:24] last week is you'll see a dramatic pickup in the tokens per second per GPU
[06:26] pickup in the tokens per second per GPU when you go from the 8-way to the 72-way
[06:29] when you go from the 8-way to the 72-way GB 200.
[06:29] Right?
[06:31] So it's curly shown but you know one thing Nvidia you know did
[06:33] you know one thing Nvidia you know did hint earlier on at NVL the idea of NVL
[06:36] hint earlier on at NVL the idea of NVL 576 right but it was never implemented
[06:39] 576 right but it was never implemented right so I think you know your argument
[06:40] right so I think you know your argument is around larger world size right and so
[06:42] is around larger world size right and so how do most of your customers um you
[06:45] how do most of your customers um you know sort of make that they use sort of
[06:46] know sort of make that they use sort of the same case so they do simulations and
[06:48] the same case so they do simulations and sort of you know why why hasn't been
[06:50] sort of you know why why hasn't been done earlier right you know it's been
[06:52] done earlier right you know it's been teased but sort of never actually follow
[06:53] teased but sort of never actually follow through with
[06:54] through with
[06:56] right so I think same if you attended the presentations today you're going to
[06:58] the presentations today you're going to find that it is starting to get deployed
[07:02] find that it is starting to get deployed as to there are many issues you know
[07:04] as to there are many issues you know there there's it all boils down to cost
[07:06] there there's it all boils down to cost right tookomics and cost uh there's
[07:09] right tookomics and cost uh there's reliability issues associated with it
[07:10] reliability issues associated with it there's high volume manufacturing issues
[07:13] there's high volume manufacturing issues associated with it so there are a bunch.
[07:14] associated with it so there are a bunch of things that have constrained CPU in the past.
[07:16] of things that have constrained CPU in the past uh what changed things as most people know here was the announcement at the GDC.
[07:19] the past uh what changed things as most people know here was the announcement at the GDC and to me that is a high volume.
[07:21] people know here was the announcement at the GDC and to me that is a high volume manufacturing foundry getting involved.
[07:24] the GDC and to me that is a high volume manufacturing foundry getting involved in it.
[07:26] manufacturing foundry getting involved in it so once this happens then everything becomes possible right.
[07:28] in it so once this happens then everything becomes possible right everyone starts looking at yes I can do.
[07:30] everything becomes possible right everyone starts looking at yes I can do this uh meta for example today presented.
[07:32] everyone starts looking at yes I can do this uh meta for example today presented data with you know 1 million hours 15 million hours of CPO with with uh almost zero uh reliability failure.
[07:35] data with you know 1 million hours 15 million hours of CPO with with uh almost zero uh reliability failure so that's fantastic right.
[07:37] million hours of CPO with with uh almost zero uh reliability failure so that's fantastic right that's a very good starting point.
[07:40] zero uh reliability failure so that's fantastic right that's a very good starting point that doesn't mean the problem is solved.
[07:42] fantastic right that's a very good starting point that doesn't mean the problem is solved I think a lot more data has to be gathered.
[07:43] starting point that doesn't mean the problem is solved I think a lot more data has to be gathered so why hasn't it been done in the past.
[07:46] problem is solved I think a lot more data has to be gathered so why hasn't it been done in the past uh all of the above.
[07:47] data has to be gathered so why hasn't it been done in the past uh all of the above uh is it going to be done yes you can see the momentum just attending all the conferences.
[07:50] been done in the past uh all of the above uh is it going to be done yes you can see the momentum just attending all the conferences yeah all the meetings.
[07:52] above uh is it going to be done yes you can see the momentum just attending all the conferences yeah all the meetings here.
[07:55] can see the momentum just attending all the conferences yeah all the meetings here Hey uh it's been fantastic to see the rate of adoption of CPO.
[07:56] the conferences yeah all the meetings here Hey uh it's been fantastic to see the rate of adoption of CPO and just one more comment before I end.
[07:58] here Hey uh it's been fantastic to see the rate of adoption of CPO and just one more comment before I end is CPO is not just about the electricals right.
[08:00] Hey uh it's been fantastic to see the rate of adoption of CPO and just one more comment before I end is CPO is not just about the electricals right it's about the entire ecosystem where high volume manufacturing plays a really really important part.
[08:02] the rate of adoption of CPO and just one more comment before I end is CPO is not just about the electricals right it's about the entire ecosystem where high volume manufacturing plays a really really important part so you not only.
[08:05] more comment before I end is CPO is not just about the electricals right it's about the entire ecosystem where high volume manufacturing plays a really really important part so you not only.
[08:06] just about the electricals right it's about the entire ecosystem where high volume manufacturing plays a really really important part so you not only.
[08:09] about the entire ecosystem where high volume manufacturing plays a really really important part so you not only.
[08:11] volume manufacturing plays a really really important part so you not only.
[08:13] really important part so you not only have to worry about do I have the coolest technology and the fastest technology but are you ready to manufacture this in high volume and this is where you know the OSATs the yield asse of the world come in foundaries come in right so this is absolutely critical so it's it's the entire the holist istic way of looking at the entire ecosystem to proliferate CPL.
[08:31] That's interesting.
[08:34] So the the sort of um roll out of scale out that becoming reality actually sort of paves the way for scale up to become much more much more build much more momentum.
[08:43] Eric yeah so for this question I think I echo the the so-called pinpoint pinpoint is obviously power and then cost that's the you know almost pinpoint for every product.
[08:55] So I would like to go back to I I mentioned briefly the wavegu based near fast you bring out uh the strong scale out to scale up I think scale out or the long distance application is the optical communication in have been working on the past like several decades what we are trying to push back in the day uh 20 almost 20 years ago I still work at Intel labs for sign photonics
[09:14] work at Intel labs for sign photonics what we are trying to do is to crank
[09:16] what we are trying to do is to crank more data as much as we could into a
[09:18] more data as much as we could into a single fiber like brings out the WDN and
[09:20] single fiber like brings out the WDN and then uh when we hit a bonet we bring the
[09:23] then uh when we hit a bonet we bring the coding now is painful and that make
[09:25] coding now is painful and that make totally makes sense as for long distance
[09:27] totally makes sense as for long distance communication but uh for scale up I
[09:30] communication but uh for scale up I think that's a slightly different you
[09:32] think that's a slightly different you know the range is different and I think
[09:34] know the range is different and I think the requirement for power is also
[09:36] the requirement for power is also different that's why you know we are we
[09:38] different that's why you know we are we have been actively you know and not just
[09:40] have been actively you know and not just on the device table but for the whole
[09:42] on the device table but for the whole season for the so-called Y and slow
[09:44] season for the so-called Y and slow approach so what what we try so what
[09:46] approach so what what we try so what this approach trying to do is to invert
[09:49] this approach trying to do is to invert what we have done in the past several
[09:51] what we have done in the past several decades and then Why we are doing this
[09:53] decades and then Why we are doing this because on the physics level uh the
[09:55] because on the physics level uh the power consumption actually scale super
[09:58] power consumption actually scale super linearly as you increase the data array.
[10:02] linearly as you increase the data array. So it means to it takes more power to
[10:05] So it means to it takes more power to transmit one bit if we the whole thing
[10:07] transmit one bit if we the whole thing is operate I high data array. So uh
[10:09] is operate I high data array. So uh going back the main the main you know
[10:12] going back the main the main you know value proposition is to save power. We
[10:14] value proposition is to save power. We are trying to reduce the pico per bit
[10:16] are trying to reduce the pico per bit and that sounds fantastic right?
[10:17] Why is the cost?
[10:19] The cost is the reach you know we are trying maybe we cannot uh reach
[10:22] and and and and the cost obviously
[10:24] because now we have to use many channel fiber that fiber cost will be probably
[10:28] is different from you know from long distance that we used eight or 16 16 fibers or 32 we are cutting this
[10:36] birectional.
[10:38] So the whole the whole why and slow approach I think is worth attention but of course we as I
[10:43] mentioned before we are a foundation technology company we actually active
[10:46] participate in the wavegu using germanium m as well but for the wine
[10:48] install I think that's the wine uh at least I attend several session in the
[10:52] OCP it's a active topic is actually under debate whether it's the right way
[10:57] to go but I think um the question will be for why and how slow it should be and also scalability
[11:05] Uh yeah and uh I think there are a few demonstration has already been done by u
[11:09] a few company staff and even CSP like Microsoft they have a white paper called
[11:16] Microsoft they have a white paper called you know introducing mosaic especially.
[11:18] you know introducing mosaic especially the mo uh the wine wine slow approach.
[11:21] the mo uh the wine wine slow approach and but now I think it's demo using blue.
[11:24] and but now I think it's demo using blue light micro using and silicon photo.
[11:26] light micro using and silicon photo detectors but um uh from the result the.
[11:30] detectors but um uh from the result the reach is kind of limited is within 10.
[11:33] reach is kind of limited is within 10 meter and the b.
[11:34] meter and the b micro led.
[11:35] micro led micro is yeah micro blue light blue.
[11:37] micro is yeah micro blue light blue light micro with a silicon you know the.
[11:40] light micro with a silicon you know the simos sensor the distance rather short I.
[11:43] simos sensor the distance rather short I think within 10 meter or even like less.
[11:45] think within 10 meter or even like less than 5 meter the bar is not to what we.
[11:49] than 5 meter the bar is not to what we are used to which is you know 10 tous 12.
[11:52] are used to which is you know 10 tous 12 so uh the brains a question the about.
[11:54] so uh the brains a question the about scalability of this approach and uh yeah.
[11:57] scalability of this approach and uh yeah that's what we have been working on for.
[11:59] that's what we have been working on for the whole time we want to yeah.
[12:00] the whole time we want to yeah yeah I mean we'll I think we'll we'll.
[12:02] we'll dive into um narrow and fast.
[12:04] we'll dive into um narrow and fast versus wide and slow and our sort of a.
[12:06] versus wide and slow and our sort of a couple followup questions I have and I.
[12:08] couple followup questions I have and I think we'll talk about the different.
[12:10] think we'll talk about the different vectors for scaling but I think maybe to.
[12:13] vectors for scaling but I think maybe to just kind of level set like you know.
[12:14] just kind of level set like you know could you give an example of like fast.
[12:16] could you give an example of like fast and now like you know today would be 200 gig pam 4 and then what would be like you know the modulation and the the um the data rate for just to give everyone example like
[12:25] yeah that's that's a that's a that's an interesting I wouldn't call tough
[12:29] there's a depends on play there are different answers and depends on wave flat you choose personally I would say
[12:34] you know what we have been working on is now I mentioned blue right what we are thinking now we are working on is now slightly longer wavelengths because you
[12:42] if you move to longer wavelengths the fiber loss actually is actually less and the disperation is also less and uh so maybe for example green light or even to near also near even to the near infarray and uh and so the data array I think
[12:57] yeah that actually relates to not just the physic is clear if you scale the data array per lane the peak per bit increase.
[13:07] Yeah. But uh you know we but we have to as a photonic company of 100 guys we have to interface with a legacy electrical interface.
[13:14] For example UCI we
[13:17] electrical interface.
[13:20] For example UCI we used to be 32 gab per second NRZ now it's 64 and that means we have to interface that.
[13:22] So I I think the that that depends on which Ethernet information is like 200.
[13:25] So 64 gigabits NRZ using NRZ right.
[13:29] Um yeah maybe we need to.
[13:31] Yeah maybe.
[13:32] Oh slightly less than that.
[13:34] Yeah, I would say it's a balance between the number of channels because in order to achieve the total accumulation bandwidth, right?
[13:37] That's what we care is the multiplication.
[13:39] The slower that we have, the lower power, but the more fiber we need, the more the long the higher the cost.
[13:42] So, it's a it's a it's a there's some interesting it's an interesting topic and we have been investing in this.
[13:48] Yeah.
[13:50] Not just us with our partners.
[13:52] Yeah.
[13:54] Right.
[13:55] Uh CK.
[13:56] Okay.
[13:59] Yeah. uh I think Eric talking about the power but I I want want to talk about from different angle from the performance wise right and media tech we are doing the uh a lot of A6 service for the big hyperscaler right I think one of the penpoint they're trying to do is they're trying to provide more uh computing power right so I think no
[14:17] computing power right so I think no matter Nvidia or all those hyperscaler.
[14:23] they try to squeeze more transistor into the silicon right.
[14:25] they try to put in more HBM or whatever.
[14:27] Now there's customer HBM or even people think about using memory spender to hook up more uh LPDDR.
[14:35] uh memory right I think all trying to trying to squeeze more uh computing bandwidth right.
[14:43] so in order so that's the problem uh when we trying to have generate a lot of data right.
[14:48] how you can transmit uh moving the data around is very difficult right.
[14:54] so now it's heavy still heavily depends on the copper So.
[14:56] now we have 200 and 24 g 30s and we're moving to the 400 g 30s right.
[15:02] no matter how far we we we going forward it going to hit the end right.
[15:08] like a physical the coers limitation you cannot even get out of the package right.
[15:13] so that's where the CPO come into the picture from uh our point of view right so so so I I think I
[15:19] point of view right so so so I I think I I talked to couple of customer right I
[15:21] I talked to couple of customer right I think uh the most common answer from
[15:26] think uh the most common answer from them is uh as long as copper still can
[15:30] them is uh as long as copper still can survive nobody want to move the to the uh optical because um entire ecosystem
[15:33] survive nobody want to move the to the uh optical because um entire ecosystem have to change right the packaging the
[15:36] have to change right the packaging the no matter you use wow the micro or use
[15:40] no matter you use wow the micro or use the laser entire ecosystem have to
[15:43] the laser entire ecosystem have to change that's a huge uh investment right
[15:45] change that's a huge uh investment right and so I think right now uh like CPO I
[15:49] and so I think right now uh like CPO I think still uh face a lot of challenge
[15:54] think still uh face a lot of challenge on the ecosystem like uh you mentioned
[15:57] on the ecosystem like uh you mentioned earlier right I think that's a a lot of
[16:00] earlier right I think that's a a lot of challenge we we still trying to solve
[16:02] challenge we we still trying to solve and so uh but CPO definitely is the
[16:05] and so uh but CPO definitely is the trend right because everybody know it
[16:08] trend right because everybody know it going to hit the world very soon right
[16:11] going to hit the world very soon right and I remember like uh six years ago my
[16:13] and I remember like uh six years ago my engineer told me 100 100k 30 is the end
[16:20] engineer told me 100 100k 30 is the end right.
[16:23] then three months later he said oh I can do it like 45 GB insertion loss.
[16:26] I said, "Why don't you tell me early?
[16:28] I already told my customer, right?"
[16:30] Then they told me like a 200 G is the the end, right?
[16:34] Then this time I don't believe them, right?
[16:37] I say you you got to have something in your pocket, right?
[16:41] So I think people talking about 400 G 448 is the end.
[16:45] I don't know. How do you think?
[16:46] I I think probably could be, but I I'm not saying the answer right now, right?
[16:53] in case people recalling and say oh it's 448 is not the end but so far I think 448 highly likely would be end for the uh electronics then we definitely need to move to the uh opics right and the time is getting very close right you think it's far away actually it's not right we I think we almost have the 448 ready next year so the next step is I
[17:20] Ready next year so the next step is I have to move into the copper uh going to have to move into the copper uh going to the Arctic.
[17:26] So that's a CPO uh have to uh you know happen right.
[17:29] So like uh I think Nvidia in GTC this year they announced their first generation the CPO.
[17:34] Right I think that is the uh very exciting moment we we we see the Nvidia also moving this direction because kind of as entire ecosystem to moving faster to get the solution ready.
[17:46] Yeah.
[17:48] Got it.
[17:51] Um I think next I want to ask a little bit more about the timing but I think you know you brought up a slightly different point than diet right because you said the impetus is reaching the end of copper right but dicha's point was more on you know the benefits of larger world size right.
[18:03] So you know what do you all think of what really is the impetus is it is it you know your view or kind of your view or and uh and what does that mean for the timing of CPO and then you know sort of while while you're answering that you know you know um maybe you can talk a bit about your go to market you know based on what your assessment of the impetuses and the timing.
[18:19] So maybe we go from left to
[18:21] timing. So maybe we go from left to right again.
[18:22] right again.
[18:23] Oh okay. So I think whoever wants to respond to that as a question you have different Yeah.
[18:28] Did you want to respond?
[18:30] Yeah. So okay either way.
[18:33] Right. So I I scale up world size or
[18:37] I don't think it is uh there there's a difference between what CK is saying and what I am saying.
[18:41] It it all heading towards the same direction.
[18:43] Right. This is why I'm so careful about using terms like scale up and scale out because to me it it's sort of you know going across to Iraq is kind of like within Iraq.
[18:51] uh as to timelines.
[18:55] Uh so this is actually a good question because uh we had looked at this with all our pick partners.
[19:01] uh and you know basically said what are you seeing your customers project as timelines and uh the common consensus was uh that most of the big guys are doing the products first and you've seen announcements by Nvidia and Broadcom and like CK said they're going to pipe clean the entire ecosystem leading to the m maturity of uh you know the entire OSA the testing the yield reliability stability all of that good
[19:22] reliability stability all of that good stuff and uh somewhere after that is stuff and uh somewhere after that is when the fast followers are going to when the fast followers are going to take take over and These are going to be two types, right?
[19:31] These are going to be the smaller guys who are actually still trying to leaprog Nvidia or they're going to be, you know, hyperscalers who are just trying to follow Nvidia.
[19:40] So when we looked at all this data, we came up with 2028, right?
[19:44] And this was just a number we'd come up with.
[19:46] But one of the first presentations I attended uh uh two days ago at the OCP was uh there was a presentation by light counting and they pretty much came up with the same number.
[19:54] They said 2028 is when the first CPO not the first when the CPO will start being deployed in millions right so this is a time frame of what we are looking at looking at and uh what is our uh what was your second
[20:09] go to market >> the go to market right so the go to market is uh from a GC perspective our go to market is number one partner with uh uh partner with photonix experts big players who are actually doing uh these
[20:22] players who are actually doing uh these kind of advanced photonix designs
[20:24] kind of advanced photonix designs something that is ready for high volume
[20:26] something that is ready for high volume manufacturing. Number two, partner with
[20:29] manufacturing. Number two, partner with customers preferably hyperscalers who
[20:31] customers preferably hyperscalers who are uh willing to do these proof of
[20:33] are uh willing to do these proof of concepts right with uh to to realize
[20:36] concepts right with uh to to realize their CPO vision and uh finally it'll be
[20:39] their CPO vision and uh finally it'll be to leverage UC's close relationship with
[20:41] to leverage UC's close relationship with TSMC and the entire ecosystem setup to
[20:44] TSMC and the entire ecosystem setup to try and deliver these products to the
[20:45] try and deliver these products to the market.
[20:47] market. >> Yeah, definitely. uh you talk about the
[20:50] >> Yeah, definitely. uh you talk about the two questions and especially you know
[20:52] two questions and especially you know when in the market so ASE we are the
[20:55] when in the market so ASE we are the integration everything just mentioned
[20:57] integration everything just mentioned OSAT so we are the biggest Osite
[21:00] OSAT so we are the biggest Osite definitely uh we are h we are actively
[21:03] definitely uh we are h we are actively working uh with the customers and uh
[21:07] working uh with the customers and uh diversified customers and try to put the
[21:11] diversified customers and try to put the CPO uh in the market and so time wise um
[21:16] CPO uh in the market and so time wise um in the next uh few ers uh we see the
[21:19] in the next uh few ers uh we see the path and we also see at the maturity uh
[21:23] path and we also see at the maturity uh from at the beginning from the proof of
[21:26] from at the beginning from the proof of concept to development stage and to
[21:29] concept to development stage and to qualification stage and to manufacturing
[21:31] qualification stage and to manufacturing stage. However, there is a lot of
[21:35] stage. However, there is a lot of challenges and from manufacturing u
[21:38] challenges and from manufacturing u perspective
[21:40] perspective uh we realize we need a lot of
[21:43] uh we realize we need a lot of collaborations. This is totally new from
[21:46] collaborations. This is totally new from the ecosystem from the supply chain
[21:49] the ecosystem from the supply chain system and this is involved from the
[21:52] system and this is involved from the foundry from design house and from the
[21:54] foundry from design house and from the customer from the hyperscalers and down
[21:57] customer from the hyperscalers and down to uh the testing. So testing is very
[22:01] to uh the testing. So testing is very challenge testing involve the OE optical
[22:05] challenge testing involve the OE optical testing how to do it and this is just
[22:09] testing how to do it and this is just bring up a lot of questions and
[22:11] bring up a lot of questions and discussions. So from our manufacturing
[22:14] discussions. So from our manufacturing uh point of view we see the needs of a
[22:16] uh point of view we see the needs of a CPO and we also see the needs to
[22:19] CPO and we also see the needs to overcome so many challenges and first
[22:22] overcome so many challenges and first the yield yield is very critical and it
[22:26] the yield yield is very critical and it involve optical and it involve
[22:29] involve optical and it involve electrical and also it involve the fiber
[22:32] electrical and also it involve the fiber attach. So, so this is a very new uh
[22:36] attach. So, so this is a very new uh challenges and also the opportunity for
[22:38] challenges and also the opportunity for us for ASE uh we have been heavily
[22:42] us for ASE uh we have been heavily involved the mass production readiness
[22:45] involved the mass production readiness so we see in near future will come
[22:47] so we see in near future will come through
[22:49] through >> Eric
[22:50] >> Eric >> yeah I think uh because they they're big
[22:53] >> yeah I think uh because they they're big guys we are relative young stuffs I echo
[22:56] guys we are relative young stuffs I echo I to me there's a lot of challenge for
[22:58] I to me there's a lot of challenge for us you know challenge is opportunity
[23:00] us you know challenge is opportunity right we survive because there's
[23:02] right we survive because there's challenge and for the timing. I I think
[23:04] challenge and for the timing. I I think uh as for what we're working on now, I
[23:07] uh as for what we're working on now, I think the the demo we we have some
[23:10] think the the demo we we have some demonstration already but you know like
[23:12] demonstration already but you know like I say I think 2028 or 2027 the M is
[23:14] I say I think 2028 or 2027 the M is probably a good estimation because we
[23:16] probably a good estimation because we have a demonstration you need to do a
[23:17] have a demonstration you need to do a testing and take about six six months or
[23:19] testing and take about six six months or even a year then the median is probably
[23:22] even a year then the median is probably probably right but we have been actively
[23:23] probably right but we have been actively working on like I mentioned for not just
[23:26] working on like I mentioned for not just the waveg base but also why slow and
[23:28] the waveg base but also why slow and since I'm I'm been for toning for like
[23:30] since I'm I'm been for toning for like yeah for those counts
[23:33] yeah for those counts In my grad school, I think for the the
[23:35] In my grad school, I think for the the problem is always the same. Uh the the
[23:39] problem is always the same. Uh the the like the the thermal stuff for the
[23:41] like the the thermal stuff for the waveguator
[23:43] waveguator like most of technology now being
[23:44] like most of technology now being announced using micro ring because it
[23:46] announced using micro ring because it has the smallest footprint but because
[23:48] has the smallest footprint but because it's small is a reser based cavity. So
[23:50] it's small is a reser based cavity. So um once the heat starts to generate you
[23:53] um once the heat starts to generate you need to spend extra significant power to
[23:57] need to spend extra significant power to overcome. So when say that power like
[23:59] overcome. So when say that power like this to another why you know we have a
[24:01] this to another why you know we have a lot of you know repose and a lot of
[24:04] lot of you know repose and a lot of giants are working on CPO but the timing
[24:06] giants are working on CPO but the timing is still 2028. Yeah some legitimate uh
[24:10] is still 2028. Yeah some legitimate uh reasons there. So for the I think for
[24:12] reasons there. So for the I think for the modulator way one of the things
[24:14] the modulator way one of the things might be worth some you know topics that
[24:15] might be worth some you know topics that why we have been working on it uh to
[24:17] why we have been working on it uh to minimize the power consumption for using
[24:20] minimize the power consumption for using a different type of modulator and also
[24:22] a different type of modulator and also fiber attachment like we call it
[24:23] fiber attachment like we call it coupling alignment stuff that's like the
[24:26] coupling alignment stuff that's like the eternal pain now so for nanofases I
[24:30] eternal pain now so for nanofases I think they are being overcome and for
[24:32] think they are being overcome and for the other approach that's probably
[24:34] the other approach that's probably another topic but I think they do there
[24:36] another topic but I think they do there are a few interesting idea can use uh
[24:39] are a few interesting idea can use uh what we try to achieve is uh passive
[24:42] what we try to achieve is uh passive alignment because imagine for a CPO when
[24:45] alignment because imagine for a CPO when it's really like for scale up is do
[24:46] it's really like for scale up is do drive I it's hard for me to imagine to
[24:49] drive I it's hard for me to imagine to do active alignment and put things on
[24:50] do active alignment and put things on you you need something that could just
[24:52] you you need something that could just clip and that actually works so that
[24:55] clip and that actually works so that that part I think is also been what have
[24:58] that part I think is also been what have been working on but I yeah
[25:00] been working on but I yeah >> see okay u I think for the timing right
[25:04] >> see okay u I think for the timing right um this million dollars question right
[25:06] um this million dollars question right everybody try to know so I think there's
[25:08] everybody try to know so I think there's no easy answer for this so Maybe I can
[25:11] no easy answer for this so Maybe I can uh from two different angle uh that
[25:13] uh from two different angle uh that everybody think about it, right? I think
[25:15] everybody think about it, right? I think the first still uh follow my previous
[25:17] the first still uh follow my previous statement the the total the TCO right I
[25:21] statement the the total the TCO right I think cost is matter right so as long as
[25:23] think cost is matter right so as long as the copper still can survive right I
[25:25] the copper still can survive right I think so so far still you know maybe
[25:28] think so so far still you know maybe still most cost efficiency so so so you
[25:32] still most cost efficiency so so so you can see people trying to make the copper
[25:34] can see people trying to make the copper survive right so I think you can see
[25:36] survive right so I think you can see many the CPC you know demonstration uh
[25:40] many the CPC you know demonstration uh on this year right I think that's a one
[25:42] on this year right I think that's a one example uh I think uh we're still trying
[25:45] example uh I think uh we're still trying to try to get carp survive because
[25:48] to try to get carp survive because overall the cost is uh uh should be more
[25:52] overall the cost is uh uh should be more competitive still right and then that
[25:54] competitive still right and then that will lead to uh the the second angle
[25:56] will lead to uh the the second angle right so when the the CPO or the optics
[26:00] right so when the the CPO or the optics cost can be lower or the performance the
[26:04] cost can be lower or the performance the uh divide by the cost is like you can
[26:06] uh divide by the cost is like you can gain the benefit when then can cross
[26:10] gain the benefit when then can cross over that would be uh another way to
[26:13] over that would be uh another way to think about it And to do that I think
[26:15] think about it And to do that I think the uh what Gio mentioned earlier when
[26:19] the uh what Gio mentioned earlier when the the EO can be ready right the EO can
[26:22] the the EO can be ready right the EO can in improve the ready so cost can going
[26:24] in improve the ready so cost can going down and all those testing is very very
[26:27] down and all those testing is very very important right so uh if you want to get
[26:30] important right so uh if you want to get the volume production of course you
[26:32] the volume production of course you cannot do it manually so automation is
[26:35] cannot do it manually so automation is one of the very key uh key index you can
[26:38] one of the very key uh key index you can we can watch so I think uh people are I
[26:41] we can watch so I think uh people are I think there's many many uh uh company
[26:44] think there's many many uh uh company that trying to do the automation to the
[26:46] that trying to do the automation to the alignment to the testing for the optical
[26:49] alignment to the testing for the optical engine right so I think uh those are the
[26:52] engine right so I think uh those are the key index we can we can uh observe right
[26:55] key index we can we can uh observe right if those thing is ready and volume and
[26:58] if those thing is ready and volume and price is chicken right so when once you
[27:01] price is chicken right so when once you cost is going down then you volume pick
[27:03] cost is going down then you volume pick it up if volume pick it up then the cost
[27:05] it up if volume pick it up then the cost is even lower I think maybe that's a two
[27:08] is even lower I think maybe that's a two different angle uh can offer for
[27:11] different angle uh can offer for thinking about it
[27:12] thinking about it >> got it Let's dive a bit deeper into um
[27:14] >> got it Let's dive a bit deeper into um you know how we scale CPO in the long
[27:16] you know how we scale CPO in the long run. I think there are sort of four
[27:18] run. I think there are sort of four layers uh four levers that I think about
[27:20] layers uh four levers that I think about for scaling. you know, I think about
[27:22] for scaling. you know, I think about baud rate, modulation, adding fibers,
[27:25] baud rate, modulation, adding fibers, right? Um, and lastly, WDM, right? So,
[27:29] right? Um, and lastly, WDM, right? So, you know, in your road map, like what do
[27:30] you know, in your road map, like what do you think which of these uh levers you
[27:33] you think which of these uh levers you think are going to be most important to
[27:34] think are going to be most important to scaling bandwidth? Um, and you know, can
[27:37] scaling bandwidth? Um, and you know, can you talk about for your own road map,
[27:39] you talk about for your own road map, you know, what do you think are the
[27:40] you know, what do you think are the largest scaling challenges you will have
[27:42] largest scaling challenges you will have to overcome in the future?
[27:44] to overcome in the future? >> Yeah. Um, yeah, we can start with Lehon.
[27:47] >> Yeah. Um, yeah, we can start with Lehon. Yeah.
[27:47] Yeah. >> Yeah. So um as uh O says and uh we we we
[27:52] >> Yeah. So um as uh O says and uh we we we working together with the customers on
[27:54] working together with the customers on the design requirement and regarding all
[27:56] the design requirement and regarding all this design you know you mention about
[27:58] this design you know you mention about you know the four different type of
[28:01] you know the four different type of applications and for that is very uh
[28:04] applications and for that is very uh hard to answer there's no easy qu easy
[28:07] hard to answer there's no easy qu easy answers because you have to based on the
[28:09] answers because you have to based on the applications and based on your
[28:12] applications and based on your performance requirement uh as o says we
[28:15] performance requirement uh as o says we we we co-work with our hyperscalers or
[28:19] we we co-work with our hyperscalers or customers and to tuning uh to tuning
[28:22] customers and to tuning uh to tuning what type of you know of the the models
[28:25] what type of you know of the the models like WDM for example and or what other
[28:29] like WDM for example and or what other balls but we see uh we support the
[28:31] balls but we see uh we support the flexibilities we support all these
[28:33] flexibilities we support all these different type of applications but from
[28:36] different type of applications but from what we see uh we see more and more
[28:39] what we see uh we see more and more people uh asking for you know for WDM
[28:42] people uh asking for you know for WDM type of uh uh structures or CDM CD CD
[28:47] type of uh uh structures or CDM CD CD WDM or DWDM. So that is kind of for high
[28:52] WDM or DWDM. So that is kind of for high higher higher performance requirement.
[28:54] higher higher performance requirement. But for us we have to be ready whenever
[28:56] But for us we have to be ready whenever you use different models but for us we
[28:58] you use different models but for us we have to ready for the integration
[29:00] have to ready for the integration solutions. Yeah.
[29:03] solutions. Yeah. >> Okay. So I'm going to preface my answer
[29:05] >> Okay. So I'm going to preface my answer with high volume manufacturing and what
[29:07] with high volume manufacturing and what the number one foundry in the world is
[29:10] the number one foundry in the world is supporting right now. Uh because that's
[29:12] supporting right now. Uh because that's what's going to lead to the mass scale
[29:14] what's going to lead to the mass scale adoption of CPO. So given what we know
[29:18] adoption of CPO. So given what we know now uh at this point we believe our
[29:21] now uh at this point we believe our customers and they are kind of in line
[29:22] customers and they are kind of in line with what the foundry is doing uh we are
[29:24] with what the foundry is doing uh we are looking at something like u one lambda
[29:27] looking at something like u one lambda per fiber DR4 200 gig uh PAM 4 uh as the
[29:33] per fiber DR4 200 gig uh PAM 4 uh as the process matures through the foundry
[29:35] process matures through the foundry they're looking at one to two lambdas
[29:37] they're looking at one to two lambdas they're looking at again 200 gig PAM 4
[29:39] they're looking at again 200 gig PAM 4 but CWDM and then finally uh you know
[29:43] but CWDM and then finally uh you know when the couplers the grading coupler
[29:44] when the couplers the grading coupler versus edge coupler or all those issues
[29:46] versus edge coupler or all those issues get sorted out. It's DWDM 8 to6 lambdas
[29:48] get sorted out. It's DWDM 8 to6 lambdas per fiber. You can either do 200 gig PAM
[29:51] per fiber. You can either do 200 gig PAM 4 or 64 NRC. So this is the road map
[29:54] 4 or 64 NRC. So this is the road map that uh it's not our road map, it's a
[29:57] that uh it's not our road map, it's a customer road map but it's very tightly
[29:59] customer road map but it's very tightly coupled to what is happening for the
[30:01] coupled to what is happening for the high volume manufacturing.
[30:02] high volume manufacturing. >> And how many fiber how many uh fiber
[30:04] >> And how many fiber how many uh fiber pairs would you attach?
[30:06] pairs would you attach? >> How many?
[30:07] >> How many? >> 46.
[30:08] >> 46. Got it. Yeah, I ask because this will be
[30:11] Got it. Yeah, I ask because this will be an important lead to Eric.
[30:16] >> We'll start with how many fiber pair
[30:18] >> We'll start with how many fiber pair we'll start with that Eric. How many
[30:19] we'll start with that Eric. How many fiber?
[30:20] fiber? >> It's just just simple math, right? Just
[30:21] >> It's just just simple math, right? Just for example, if we are using like 2 GB
[30:23] for example, if we are using like 2 GB per second and to reach 1.6
[30:27] per second and to reach 1.6 800, right?
[30:29] 800, right? >> 800.
[30:29] >> 800. >> Yeah.
[30:30] >> Yeah. >> Fibers.
[30:30] >> Fibers. >> Yeah.
[30:31] >> Yeah. >> There's a different fiber pitch. So
[30:32] >> There's a different fiber pitch. So fiber. So they are they actually
[30:34] fiber. So they are they actually different kind of fiber. So fiber what
[30:36] different kind of fiber. So fiber what we are used to is bra single mode is
[30:39] we are used to is bra single mode is glass based fiber the loss is really low
[30:41] glass based fiber the loss is really low and the main purpose is because the
[30:43] and the main purpose is because the original purpose for optical
[30:44] original purpose for optical communication is for long distance so we
[30:46] communication is for long distance so we need really low and the the cost is
[30:48] need really low and the the cost is there but there other fiber for example
[30:50] there but there other fiber for example the one used in the medical examination
[30:53] the one used in the medical examination the endoscope that's for imaging purpose
[30:55] the endoscope that's for imaging purpose the fiber is actually made of plastic
[30:57] the fiber is actually made of plastic and maybe using some neurons the loss is
[31:00] and maybe using some neurons the loss is significantly higher than glass that's
[31:03] significantly higher than glass that's why it brings to it's all about the
[31:05] why it brings to it's all about the application. If what we want to transfer
[31:07] application. If what we want to transfer is not maybe five meter is too short but
[31:11] is not maybe five meter is too short but what we want if it's 30 meter 50 meter
[31:13] what we want if it's 30 meter 50 meter then maybe this kind of fiber is good
[31:16] then maybe this kind of fiber is good enough that's why I go back to what we
[31:18] enough that's why I go back to what we are working on because the fiber loss is
[31:20] are working on because the fiber loss is significant so we need uh from system we
[31:24] significant so we need uh from system we need a really high sensitive detector
[31:26] need a really high sensitive detector such that we can minimize the required
[31:29] such that we can minimize the required emitter power no matter it's micro
[31:32] emitter power no matter it's micro resonant cavity based micro or just
[31:34] resonant cavity based micro or just vixel at low
[31:36] vixel at low Yeah. So the the fiber number can be
[31:38] Yeah. So the the fiber number can be drastically I think the audience is very
[31:40] drastically I think the audience is very familiar with the you know narrow phase.
[31:42] familiar with the you know narrow phase. Yeah. It's just you know 200 200 gig
[31:44] Yeah. It's just you know 200 200 gig then you need a for 126 and birectional
[31:48] then you need a for 126 and birectional 60. But I I think that's kind of
[31:50] 60. But I I think that's kind of different approach. So going all I echo
[31:53] different approach. So going all I echo uh c it's all going back to cost. So the
[31:55] uh c it's all going back to cost. So the wire slow project from our we are
[31:58] wire slow project from our we are probably from cost perspective we might
[32:01] probably from cost perspective we might use a less you know higher loss fiber
[32:03] use a less you know higher loss fiber with more channel the the the core is
[32:06] with more channel the the the core is really tiny if you you can check for
[32:08] really tiny if you you can check for it's not it's made for medical so for
[32:10] it's not it's made for medical so for imaging purpose but anyway that's the
[32:12] imaging purpose but anyway that's the other approach we we are pursuing so uh
[32:15] other approach we we are pursuing so uh we have shown many interesting
[32:16] we have shown many interesting demonstration with our customer like
[32:18] demonstration with our customer like using our detector which is you know uh
[32:21] using our detector which is you know uh really high I would say our key we have
[32:24] really high I would say our key we have really high low dark current that means
[32:25] really high low dark current that means our sensitivities is is like a 10 dB
[32:28] our sensitivities is is like a 10 dB above what others can provide even at
[32:31] above what others can provide even at 125 degrees CC and uh yeah for certain
[32:33] 125 degrees CC and uh yeah for certain application right 30 m is we think it's
[32:36] application right 30 m is we think it's good enough yeah
[32:37] good enough yeah >> and so I think that the challenge like
[32:39] >> and so I think that the challenge like my my guess the scaling challenge would
[32:41] my my guess the scaling challenge would possibly be the the fiber pitch right
[32:43] possibly be the the fiber pitch right because I think the the industry is
[32:45] because I think the the industry is converging around 127 microns right and
[32:47] converging around 127 microns right and so um you know for your sort of idea of
[32:50] so um you know for your sort of idea of you know how wide and slow would it look
[32:51] you know how wide and slow would it look like what kind of pitch would you need
[32:53] like what kind of pitch would you need to have
[32:53] to have >> yeah there is paper like you know uh
[32:56] >> yeah there is paper like you know uh some there's other company working list
[32:58] some there's other company working list they have their pitch on top yeah it's
[33:00] they have their pitch on top yeah it's kind of puppy data but as far as I know
[33:02] kind of puppy data but as far as I know we are talk with fiber company with our
[33:04] we are talk with fiber company with our partners they can go to pretty small
[33:06] partners they can go to pretty small pitch like less than 10 micron
[33:08] pitch like less than 10 micron >> less than 10 microns got even though I
[33:10] >> less than 10 microns got even though I think the core can be like eight eight
[33:12] think the core can be like eight eight microns right for DR at least
[33:14] microns right for DR at least >> yeah yeah but it's like again lie is not
[33:16] >> yeah yeah but it's like again lie is not a glass fiber so it's a different it's a
[33:19] a glass fiber so it's a different it's a different type lousier fiber
[33:21] different type lousier fiber >> I see I see
[33:22] >> I see I see >> but I I shouldn't say I didn't any but
[33:27] >> but I I shouldn't say I didn't any but you know yeah so is this about the whole
[33:29] you know yeah so is this about the whole you know cost and performance stuff
[33:31] you know cost and performance stuff >> and I guess
[33:31] >> and I guess >> but it's cheaper so I see yeah if
[33:33] >> but it's cheaper so I see yeah if >> and I think it's as much cost as it is
[33:36] >> and I think it's as much cost as it is power right because you know wide and
[33:37] power right because you know wide and still another reason is you know when
[33:39] still another reason is you know when you kind of go to lower bit rate in your
[33:41] you kind of go to lower bit rate in your NRZ then it's power as well right so
[33:43] NRZ then it's power as well right so that would be sort of a point that would
[33:45] that would be sort of a point that would argue maybe against sticking with pam 4
[33:48] argue maybe against sticking with pam 4 200 gig or 400 gig as a case
[33:49] 200 gig or 400 gig as a case >> yeah that's the let's see you mention
[33:51] >> yeah that's the let's see you mention the main bottom I think pam for itself
[33:53] the main bottom I think pam for itself is uh is is from electrical uh you know
[33:57] is uh is is from electrical uh you know perspective you need to incre need this
[33:59] perspective you need to incre need this DSP to you know to overcome the channel
[34:01] DSP to you know to overcome the channel loss basically and but for from optical
[34:05] loss basically and but for from optical or folks we can work with P4 as well
[34:06] or folks we can work with P4 as well this is a linear photo detector and
[34:08] this is a linear photo detector and linear type of emitter as well but uh if
[34:12] linear type of emitter as well but uh if going back to the the power if we want
[34:14] going back to the the power if we want to achieve really low power I think
[34:16] to achieve really low power I think might be an easier solution if we can do
[34:18] might be an easier solution if we can do that just like car is easier solution if
[34:20] that just like car is easier solution if we can deliver yeah Yeah. Um let's hear
[34:23] we can deliver yeah Yeah. Um let's hear from CK and then um maybe open up to the
[34:25] from CK and then um maybe open up to the audience after that.
[34:26] audience after that. >> Okay. Yeah. So uh I I think from uh our
[34:30] >> Okay. Yeah. So uh I I think from uh our point of view uh those different
[34:32] point of view uh those different architecture may exist because uh all
[34:35] architecture may exist because uh all those uh hybrid scaler they have
[34:37] those uh hybrid scaler they have different architecture in their data
[34:39] different architecture in their data center right so I think from the
[34:41] center right so I think from the traditional lift spine those traditional
[34:44] traditional lift spine those traditional uh mesh networking or like Google they
[34:47] uh mesh networking or like Google they have the uh pointto-point this kind uh
[34:50] have the uh pointto-point this kind uh uh architecture actually they are
[34:52] uh architecture actually they are requiring different u uh uh scheme right
[34:56] requiring different u uh uh scheme right like a mesh network uh multi- lambda per
[35:00] like a mesh network uh multi- lambda per fiber probably not really useful for
[35:02] fiber probably not really useful for them right so the the DR4 like 200 g or
[35:05] them right so the the DR4 like 200 g or 400 g uh uh per lambda maybe that's the
[35:10] 400 g uh uh per lambda maybe that's the the technology they are looking for
[35:11] the technology they are looking for right but like for for different they
[35:14] right but like for for different they might may looking for like a CWDM or DWM
[35:16] might may looking for like a CWDM or DWM trying to squeeze as many as uh uh
[35:19] trying to squeeze as many as uh uh lambda into the same fiber try to
[35:21] lambda into the same fiber try to squeeze more bandwidth right so uh like
[35:25] squeeze more bandwidth right so uh like so media tech we as a as provider to
[35:28] so media tech we as a as provider to serve all the those hypers scale we
[35:30] serve all the those hypers scale we pretty much we will provide all all kind
[35:33] pretty much we will provide all all kind of a solution both of the solution you
[35:35] of a solution both of the solution you know so that's why uh we develop our own
[35:38] know so that's why uh we develop our own pic and EIC and we try to provide the
[35:41] pic and EIC and we try to provide the solution but but again I think I want to
[35:43] solution but but again I think I want to echo uh yong earlier right this is not
[35:47] echo uh yong earlier right this is not the uh it's very diff complicated system
[35:50] the uh it's very diff complicated system right so it's not only one company can
[35:52] right so it's not only one company can do it right so we still need a lot of
[35:54] do it right so we still need a lot of partner even Um even we have our own PIC
[35:58] partner even Um even we have our own PIC right we still need to work with many
[36:00] right we still need to work with many you know partner because there's too
[36:03] you know partner because there's too many uh uh protocol or too many
[36:06] many uh uh protocol or too many architecture need to fulfill right so so
[36:09] architecture need to fulfill right so so it's not very simple uh solution can be
[36:12] it's not very simple uh solution can be fulfill all those requirement because
[36:14] fulfill all those requirement because when the the time is I think the
[36:17] when the the time is I think the technology is still evolving quickly
[36:19] technology is still evolving quickly right so people all those hyperscalers
[36:22] right so people all those hyperscalers are still trying to find the most
[36:24] are still trying to find the most efficient way to to squeeze more
[36:26] efficient way to to squeeze more bandwidth. Yeah.
[36:28] bandwidth. Yeah. >> Great.
[36:28] >> Great. >> Uh any questions from the audience? You
[36:30] >> Uh any questions from the audience? You can come up. Can you come up to the
[36:32] can come up. Can you come up to the microphone?
[36:34] microphone? Yeah.
[36:37] Yeah. >> Hello. Yes. But uh actually I have a two
[36:39] >> Hello. Yes. But uh actually I have a two question. The first question is like you
[36:41] question. The first question is like you know regard what is the biggest
[36:44] know regard what is the biggest dimension and your team are working on
[36:47] dimension and your team are working on that. The second question is regarding
[36:49] that. The second question is regarding manufacturing. So from this assembly
[36:52] manufacturing. So from this assembly point and the testing point. So what the
[36:54] point and the testing point. So what the most is the challenge just like in a
[36:56] most is the challenge just like in a process is there process is there okay
[36:59] process is there process is there okay regarding the equipment thank you
[37:04] >> so uh I may you know just mention about
[37:06] >> so uh I may you know just mention about the manufacturing yeah so okay so there
[37:09] the manufacturing yeah so okay so there you just mentioned about the challenges
[37:11] you just mentioned about the challenges uh we have so many challenges but point
[37:14] uh we have so many challenges but point of the new things for the CPO definitely
[37:18] of the new things for the CPO definitely is optical engine so optical engine
[37:21] is optical engine so optical engine there's deolve uh two things and just
[37:24] there's deolve uh two things and just like you know Eric or here they talk
[37:26] like you know Eric or here they talk about besides you know you have the OE
[37:30] about besides you know you have the OE like you know from the foundry because
[37:32] like you know from the foundry because foundry so you you can see how many
[37:34] foundry so you you can see how many foundry they can do the optical engine
[37:36] foundry they can do the optical engine right so you can use your finger to
[37:39] right so you can use your finger to point it out
[37:41] point it out so so that is the from the foundry but
[37:43] so so that is the from the foundry but uh come to us we do everything
[37:46] uh come to us we do everything integration so we do the OE and we do
[37:49] integration so we do the OE and we do the ASC and we also do the assembling on
[37:53] the ASC and we also do the assembling on the laser dies and on the fiberach.
[37:56] the laser dies and on the fiberach. So there's manufacturing technology in
[37:59] So there's manufacturing technology in uh innovation we need. We need a new
[38:02] uh innovation we need. We need a new technology people talk about for example
[38:05] technology people talk about for example 2.5D silicure that is being popular used
[38:08] 2.5D silicure that is being popular used right now but that is not only
[38:11] right now but that is not only solutions. So from the packaging
[38:14] solutions. So from the packaging technology we need innovation and
[38:16] technology we need innovation and secondly the test we want to receive all
[38:19] secondly the test we want to receive all the good die people talk about the
[38:21] the good die people talk about the chibulus and they already the good non
[38:24] chibulus and they already the good non good die is the concept being become
[38:27] good die is the concept being become blur they say the non bad die so then
[38:30] blur they say the non bad die so then that is on the as die or on the logical
[38:33] that is on the as die or on the logical die but on the optical die OE engine so
[38:37] die but on the optical die OE engine so we will receive the OE and who will in
[38:39] we will receive the OE and who will in charge who call to test
[38:42] charge who call to test And there are so many things and being
[38:45] And there are so many things and being developed but still not be mutual enough
[38:49] developed but still not be mutual enough thin D or thick whole D to do the OE die
[38:52] thin D or thick whole D to do the OE die test and after that you do all the fiber
[38:55] test and after that you do all the fiber and the driver and who do the system
[38:58] and the driver and who do the system level test is not we do not have the
[39:01] level test is not we do not have the total solution yet. So that all involve
[39:04] total solution yet. So that all involve the yield. So that is very challenging
[39:07] the yield. So that is very challenging from the manufacturing and also you talk
[39:10] from the manufacturing and also you talk about the two the two is not fully uh
[39:14] about the two the two is not fully uh mutual yet especially for the automation
[39:17] mutual yet especially for the automation we need automation but uh for the CPO
[39:20] we need automation but uh for the CPO automation on the tooling we still need
[39:23] automation on the tooling we still need the two suppliers to develop new for the
[39:27] the two suppliers to develop new for the mass production requirement. Yeah,
[39:30] mass production requirement. Yeah, >> may maybe we've got two more minutes.
[39:32] >> may maybe we've got two more minutes. Maybe we can take one more question from
[39:33] Maybe we can take one more question from the audience.
[39:35] the audience. >> So I'm sorry just uh the first question
[39:37] >> So I'm sorry just uh the first question just like what is the biggest dimension
[39:40] just like what is the biggest dimension just like a CPU and you guys working on
[39:42] just like a CPU and you guys working on that?
[39:42] that? >> What do you mean biggest dimension?
[39:44] >> What do you mean biggest dimension? >> Uh regarding regarding the CPU the size
[39:47] >> Uh regarding regarding the CPU the size >> dimension of the CPU the the pick or
[39:49] >> dimension of the CPU the the pick or >> yeah the
[39:52] >> yeah the size the size of the chip. No
[39:55] size the size of the chip. No >> right. So
[39:56] >> right. So >> I don't know the exact size but we are
[39:59] >> I don't know the exact size but we are looking at putting optical engines on a
[40:01] looking at putting optical engines on a 5 1/2x reticle and uh I think we can put
[40:04] 5 1/2x reticle and uh I think we can put maybe four. We are still investigating
[40:06] maybe four. We are still investigating how many can actually be put in onto a
[40:08] how many can actually be put in onto a actual 5 and 1/2 x reticle. So four on
[40:11] actual 5 and 1/2 x reticle. So four on each side maybe five on each side.
[40:14] each side maybe five on each side. >> Yeah the land just
[40:16] >> Yeah the land just >> there's some good demos actually on the
[40:18] >> there's some good demos actually on the trade floor. Yeah. Give it to you
[40:19] trade floor. Yeah. Give it to you offline actually. Yeah there's some good
[40:20] offline actually. Yeah there's some good demos on the trade floor. You can see
[40:21] demos on the trade floor. You can see all the uh I think there's one. Yeah.
[40:24] all the uh I think there's one. Yeah. Okay. Next question.
[40:26] Okay. Next question. >> Last one though. We have one minute.
[40:28] >> Last one though. We have one minute. >> Okay.
[40:29] >> Okay. >> Several highlighted manufacturability is
[40:32] >> Several highlighted manufacturability is a key thing and I think you know the
[40:34] a key thing and I think you know the silicon integration I mean there's a lot
[40:36] silicon integration I mean there's a lot of challenges but we see those being
[40:38] of challenges but we see those being overcome. But if I think about uh the
[40:41] overcome. But if I think about uh the handling of a package with CPO whether
[40:44] handling of a package with CPO whether it be through assembly or through test
[40:47] it be through assembly or through test the fibers become a huge obstacle. Um so
[40:51] the fibers become a huge obstacle. Um so what just an assessment what's the state
[40:53] what just an assessment what's the state of the industry in terms of readiness
[40:55] of the industry in terms of readiness for attaching fiber last or using uh
[41:00] for attaching fiber last or using uh pluggable fiber attached for tests or
[41:02] pluggable fiber attached for tests or what a way to automate this process
[41:04] what a way to automate this process because clearly you can't go through
[41:06] because clearly you can't go through reflow with fibers hanging off these
[41:08] reflow with fibers hanging off these packages. Automating tests with fibers
[41:11] packages. Automating tests with fibers hanging off is not going to be easy.
[41:16] >> Manufacturability all yours.
[41:19] >> Manufacturability all yours. No idea. I have no idea.
[41:22] No idea. I have no idea. >> Okay. So yeah, Aio know us very well and
[41:26] >> Okay. So yeah, Aio know us very well and uh so regarding how to handling uh you
[41:28] uh so regarding how to handling uh you know the fiber attach actually um in the
[41:31] know the fiber attach actually um in the aso sides we team up uh with the fiber
[41:35] aso sides we team up uh with the fiber you know majority on the CM as you know
[41:38] you know majority on the CM as you know CM size is handling majority on the
[41:41] CM size is handling majority on the fiber fiber really fiber attach but on
[41:44] fiber fiber really fiber attach but on we dealing with you know how the fiber
[41:47] we dealing with you know how the fiber alignment you know passive alignment
[41:49] alignment you know passive alignment active alignment so currently there is
[41:52] active alignment so currently there is the two as you know one is we call the
[41:55] the two as you know one is we call the detachable so detachable kind of working
[41:59] detachable so detachable kind of working better because you can reduce all this
[42:02] better because you can reduce all this directly fiber attach so we we working
[42:05] directly fiber attach so we we working on detachable we see that is recently in
[42:08] on detachable we see that is recently in the in just one two years just developed
[42:11] the in just one two years just developed we think that is maybe the direction for
[42:14] we think that is maybe the direction for all sides to handle but if you talk
[42:16] all sides to handle but if you talk about all this direct attach we still
[42:19] about all this direct attach we still need the working with the CM of the
[42:21] need the working with the CM of the fiber company that is not our current is
[42:25] fiber company that is not our current is not our
[42:26] not our >> I guess that's where I was going is
[42:28] >> I guess that's where I was going is what's the maturity of the providers of
[42:30] what's the maturity of the providers of these detachable
[42:32] these detachable >> solutions right whether they're sockets
[42:34] >> solutions right whether they're sockets or whatever you want to call it
[42:36] or whatever you want to call it >> okay that's a very good question and I
[42:38] >> okay that's a very good question and I attended the OFC uh this year still
[42:41] attended the OFC uh this year still under development is not mutual mutual
[42:43] under development is not mutual mutual for mass production
[42:45] for mass production >> probably scaling
[42:46] >> probably scaling >> scaling yes
[42:47] >> scaling yes >> I think we'll take that offline we have
[42:49] >> I think we'll take that offline we have to uh end the panel 501. But okay,
[42:52] to uh end the panel 501. But okay, thanks so much everyone.