The physical limits of copper are forcing a shift to optical interconnects in AI data centers. Austin sits down with GlobalFoundries' Thomas Barber to unpack why GF thinks it can lead that transition twice over: once with its long-running silicon photonics platform, and again with the specialty Silicon Germanium process the industry needs to drive it. "The enemy to me right now is copper. I'm trying to beat copper, right? If TSMC wins and we win, that's great because we're both displacing copper. And until all the copper is gone, there's plenty of market to go around." - Thomas Barber, GlobalFoundries Key Takeaways: - Copper's usable range halves every time the data rate doubles, and 200 Gbps/lane inside a rack-scale AI cluster is already past the point copper can handle. - CPO's real win isn't speed, it's power: saving 20-25 pJ/bit frees part of a data center's fixed 50-100 MW budget to go toward compute instead of moving bits. - GlobalFoundries leads photonics revenue for an unglamorous reason: it moved to 300mm wafers early, which yields 2.25x more die per wafer than the 200mm lines rivals still run. - CPO can end up more reliable than the pluggables it's replacing, not less, because it deletes the physical plug connector, and dust at that connector is the leading cause of field failures. - The OCI MSA picks NRZ over faster PAM4, deliberately going wide and slow, because NRZ's native bit error rate is a million times lower, which simplifies the receiver and cuts power. - GlobalFoundries stacks two specialty processes into one edge: micro-mirror couplers on the photonic side, and Silicon Germanium transistors hitting 350-400 GHz on the electrical side driving them. - The real competitor for silicon photonics isn't another foundry, it's copper itself - and that market is big enough that GlobalFoundries and TSMC can both win without taking share from each other. Chapters: 0:00 GlobalFoundries in Photonics 1:21 GF's Photonics Strategy 3:02 GF's Market Leadership 5:06 300mm Wafer Advantage 6:55 Copper's Range Limits 9:18 Pluggable to CPO 13:48 CPO Reliability 19:22 OCI MSA Explained 26:20 GF's Scale Platform 29:04 Micro Mirror Technology 31:32 Photonics vs. Copper 38:25 Silicon Germanium Advantage Follow Chipstrat: Newsletter: https://www.chipstrat.com X: https://x.com/chipstrat Follow Vik: Newsletter: https://www.viksnewsletter.com/ X: https://x.com/vikramskr Follow Semi Doped: Get more of Austin and Vik daily, free: https://daily.semidoped.com/
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
SPEAKER_00Hello everyone. Today we have a special guest, Tom Barber, VP Communications Infrastructure and Data Center at Global Foundries. Hey Tom. Good morning.
SPEAKER_01How are you?
SPEAKER_00Good, good. Thanks for coming. So today I want to talk about Global Foundries and your photonics and optics, sort of background and experience as a company. I don't think most people think of Global Foundries when they hear CPO or when they're thinking about photonics. So I thought you'd be the perfect person to talk to on this topic. So let's jump in. So, you know, I think right away when I heard of Global Foundries in this space, I thought, wait, Global Foundries does photonics.
SPEAKER_00So, you know, most people, probably like me, know Global Foundries as the foundry that was in the running at the leading edge for a while and then stepped back and started focusing really on uh foundational nodes. Um but what people like myself don't know is that you actually do have a background uh in the business of moving data with light. And you have for well over a decade. Um I think it traces back to the IBM microelectronics business. So maybe start very high level for us. Like, what does Global Foundries even do in photonics?
SPEAKER_00How far back does it go? Just educate us.
SPEAKER_01Sure. Yeah, and I think you know, Global Foundries focus is really on what we call essential silicon, right? And that's the foundation of all our electronic systems, you know, including data centers, right? It's power, it's communications, you know, it's memory, it's it's all the things that are essential to making these communications, these electronic systems work. And Silicon Photonics perfectly aligns with that strategy. You know, it's very differentiated, it's very complex technically, um, and it's in a very challenging market, right? So there's a great opportunity for to differentiate.
SPEAKER_01Um, for those people that don't know, you know, Global Foundries, we were born out of AMD. Um, we were the AMD foundries that were spun off. But over time, we acquired both chartered semiconductor, which had a silicon photonics effort, and we acquired IBM Microelectronics, which has a long history in silicon photonics. The effort at uh IBM Microelectronics actually started in the mid-2000s with you know doing very basic integration of waveguides, modulators, and photodiodes into CMOS technologies.
SPEAKER_01In the mid-2010s, so after about 10 years of effort, that turned into real products shipping, you know, 25 gig per lane type of products shipping in mass volume. Um the volumes were fairly small at the time because it was really only used for long-range point-to-point communication. But we've continued to invest, and now we're on our third generation of integrated silicon photonics, and that's really what's being rolled out in mass scale in the data centers today.
SPEAKER_00Hmm. So interesting. Okay, so the history traces back, you got AMD and acquisitions of chartered semi and IBM microelectronics. That's super interesting. And yeah, it sounds like you've been in this uh long-distance optical communication space. So I did not know that. Now, what also surprised me is I'd heard recently that Global Foundry claims to be the largest pure play silicon photonics foundry by revenue. And again, I think people think of um TSMC Coupe or they think of Tower Semi.
SPEAKER_00And again, that's kind of like a shocking fact. So can you maybe tell us a little bit more about like, is that true? And tell us a little bit more about the business.
SPEAKER_01Yeah, so that is true. I think that's really, you know, we have been working with key customers for a very long time, right? And growing as the business grows. Um, and that's been key to us to make sure that we have those strategic customers in place that are the market leaders. The second piece of this is we made a very early investment in moving silicon botonics to 300 millimeter foundries, right? We put it into our facility in Malta, New York. Um, and that's been huge for us because it has allowed us to significantly expand our capacity, and that's what the market's looking for as much as anything right now, is there everybody's just desperate for more capacity, and we are available to provide that capacity.
SPEAKER_01As a part of that, we added the AMF team in Singapore, which brought in a full range of very highly differentiated products that were actually orthogonal to what GF was doing. GF's primary focus was on short range and data center products, where AMF was doing a lot of business in long range. And so that helped us round out our portfolio and added a complementary set of very high value products. And the combination of the GF Silicon Photonics business plus the AMF business is what made us number one by revenue.
SPEAKER_00Ah, okay, okay. So there was the short range play and the long range via the acquisition. Um, one interesting thing, going back a little bit, you said moving to 300 millimeter um wafers. Can you explain that more? Like, was that coming from 200 millimeters to 300 millimeters? And why is that such a big deal to unlock more capacity?
SPEAKER_01Yeah. So yes, most silicon photonics today is actually still produced on 200 millimeter wafers. That's where most of our competitors are. When you go to a 300 millimeter wafer, you know, it's it's the radius squared, right? So you go up by 50% of radius, you square that, you know, you get 2.25 times as many dye per wafer, right? So running the same number of wafers, you get more than twice as many units coming through. And that's really important as we scale. You know, photonics, even going back five years, was a relatively small business, right?
SPEAKER_01For two reasons. One is, you know, there were other alternatives, indium phosphide and EML, that were very competitive at 50 and 100 gigabits per second. Plus, you know, the overall market was just significantly smaller. The data center piece had not really taken off yet. But the data center piece taking off and 200 gigabits per second being the node that's really taking off, that's the optimal place for silicon photonics. And that's why we've seen an explosion in silicon photonics. So having that extra volume available via 300 millimeter wafers is really helpful.
SPEAKER_00Yes. Okay, so let's talk more about this 200 gigabits per second and sort of the opportunity right now for uh optics and and maybe let's talk like scale up. Um so I know you know early rack scale AI systems, the switch was at the top of the rack. You had like eight GPUs interconnected and nodes, and they were connected to a switch at the top of the rack. But um, as these racks are getting denser and we've moved to 200 um gigabits per lane speed, now all of a sudden um I think it's difficult to get the I know the switch has moved to the middle of the rack because it's harder to reach from top to bottom.
SPEAKER_00So, like unpack the problem that the industry is trying to overcome right now.
SPEAKER_01Yeah. So it all comes down to the fundamental limitation of copper, which is range. Right? So at 100 gigabits per second, a direct attached copper cable, so no intelligence, just a dumb cable, could span about two meters, which is basically the height of a rack, right? So you could go, you know, any connection within a rack, you could do with copper without a problem. When you get to 200 gigabits per second, that range drops to about one meter, which is why when you look at like an NVL72, they move the switches to the middle of the rack, because that's kind of as far as they can go realistically.
SPEAKER_01When you go to 400 gigabits per second, now the range is half a meter, and you you you're not gonna quarter the size of the rack, right? So you have to go to something faster or something, you know, something with more range, which is optical. That's really kind of the high speed stuff, and that's within a rack. Now, right now, most of the scale-up networks are within a rack. So an NVL72 is one rack, the helios that was just announced by AMD is a single rack, but you really want to go to multiple racks. And when you go to multiple racks, now you're talking about spanning tens or even hundreds of meters, and that's when you absolutely have to switch to optical.
SPEAKER_01You cannot do that with copper, not at the speeds we want to do.
SPEAKER_00Okay, so for scale up, you're saying that copper's getting shorter and shorter. And so one way to move forward is to just try to move the switches closer and closer to the GPUs, but ultimately that's gonna run out of steam. So even within the rack, you might need optics. But then, of course, the question is why stop with one rack? Why not have a scale-up domain that could stretch into neighboring racks or or four racks side by side? And so, of course, now it was about like shrinking, but then it's about expanding as well.
SPEAKER_00And you're saying that only optics can pull that off.
SPEAKER_01Yeah, if you if you look at what Google has done and even what Huawei has done with their scale-up networks, they've gone beyond a single rack. So instead of being limited to you know 72 or 144 GPUs, depending on how you you count them, you know, they're doing 300, 500, 1000 GPUs across six, eight, 10 racks, right? So they're really expanding the size of their scale-up networks using optical communication to span multiple racks.
SPEAKER_00Gotcha. Okay, so then walk me through are we talking about like pluggables for the optics or NPO or CPO? And and where does Global Foundries come in?
SPEAKER_01Right. So it can be any of those options. What we see is a trend toward from pluggables first going toward NPO and then toward CPO. And the reason is really cost and power and density. So when you're at the edge of the rack with a pluggable, you're looking at about 35 dB of loss from the edge of the rack to the GPU or switch ASIC in the middle of the server. Right? That loss has to be compensated for by a you know kind of expensive and pretty power-hungry uh DSP in the pluggable.
SPEAKER_01When you move to an NPO system, now you've dropped your loss from about 35 dB to somewhere between 15 and 20 dB. That allows you to significantly reduce the amount of compensation you need to do on the electrical side or even eliminate it.
unknownRight?
SPEAKER_01So if you have enough link budget end-to-end, that you can handle the extra 15 dB of loss on either end due to the transition from the NPO to the core processor, then you can get away with not having a read timer in the NPO, which is what most people are doing. They're doing what's called linear. And so that removes that DSP from the system, which saves a lot of cost and it saves a lot of power. But you still have a fairly bulky package for this NPO.
SPEAKER_01And what you want to do, and you still have a lot of loss in the system, right? There's still 15 dB of loss that needs to be compensated for. If you continue to co-package the optics and you move that transition from electrical to optical into the package, now you've dropped your loss from you know 15 dB down to like 6 dB. And that really helps you both from the link budget, obviously, because you've gotten yourself an extra 20 dB, but that also allows you to lower the power of your CERTEs significantly.
SPEAKER_01So, what we're trying to do is we're trying to minimize the number of picojoules per bit that you need for the communication. So when you're a pluggable, you're talking about, you know, a fully retimed pluggable is somewhere around 20 to 25 picojoules per bit. When you go to NPO, you can go to linear and get rid of the DSP, which does a lot. It gets you down to maybe 10 picojoules per bit, but you still have a pretty powerful CERT to overcome that 15 dB of loss between the coprocessor, core processor, and the uh NPO.
SPEAKER_01When you go onto the package with the GPU or the or the ASIC, now you're talking about 3 dB of loss, and now you can go to the lowest power CERTs, which allows you to get down to less than five picajoules per bit. Now that's really critical because all of these systems are thermally limited. So every picojoule of power that goes into the communications is one you can't use compute. So what you want to do is minimize the amount of power that you use for the communications between the GPUs so you can run the GPUs as fast as possible.
SPEAKER_00Sure. Okay, so if we orient around power, the goal is to in a gift given a fixed power budget in this data center, it's only got 50 megawatts or 100 megawatts, whatever. We want to use as much power as possible for compute and as little as power as possible for communication, sending data around. Now, talk us through again why, kind of zooming back out. Um, you talked about pluggables. Why do they have so much loss in them? You talked about they have a DSP, but like where does the loss come from?
SPEAKER_00Because I think people, it's not in the optics, right? It's like the electrical trace.
SPEAKER_01Yeah, it it what it is, it's it's the it's the electrical traces through the PCBs from the core processor, you know, a couple hundred centimeters to the edge of the server, right? And so those are very thin wires. Um, and when you're talking about, you know, a 200 gigabit per second PAM4 signal has about 55 gigahertz of bandwidth. Right. So you're trying to run a 55 gigahertz signal through a very long trace on a PCB, and you end up with a lot of loss.
SPEAKER_00So then a step in the right direction, uh, NPO, near pluggable optics, you're ultimately shortening that trace, and then finally co-package optics, maybe you're getting rid of it entirely. But can you walk us it? Can you walk us through the trade-offs of, you know, why doesn't why doesn't ever I know that people are talking about using NPO as a stepping stone to CPO. So can you talk us through like why not just jump to CPO if it's the most power efficient?
SPEAKER_01Right. So the concern with moving away from pluggables is reliability. Right? So right now, optical communications in general is seen as a reliability risk. And so the data center customers are very comfortable with it being pluggables because if something goes wrong, they just swap the pluggable, right? Just take one out, put a new one in, and move on. They don't always do root cause to ensure that the pluggable is really the problem.
SPEAKER_01But if you swap out the pluggable and the problem goes away, then you're fine, you keep going. Moving to near-packaged optics is a half step in the right direction. Those solutions will probably be socketed, so they'll still be relatively easy to maintain. You'll still have to open you'll have to open the server and swap it out physically, but it's still something that can be swapped in place without completely replacing the GPU. When you go to co-packaged optics, now you're talking about something that is permanently attached inside the package of the GPU.
SPEAKER_01So the reliability has to be extremely high. So data center customers, Meta in particular, are going through the effort to validate the reliability of those solutions right now. And Meta's released a couple versions of a study that they've had running for a long time with you know 50,000, I think it is, GPUs in a network using co-package optics. And what they found is the reliability is actually higher than it is with pluggables. And they found no link flap, which is the primary cause of concern in optical communications across their entire testing range, which I think is up to 50 million hours or something like that right now.
SPEAKER_01So, you know, they're they're doing the work and becoming more comfortable with the reliability, but they're not going to deploy it at scale until they're absolutely certain that it's a reliable solution.
SPEAKER_00Fascinating. So tell us why why do people why are people expecting that the reliability is not great for CPO and then finding out from meta, like, oh, actually it's not as bad as you think. I mean, is this about like, is the laser, is it an external laser? And the concern was that the laser's sitting near a hot chip, but that's not a problem? Or or what is the problem that people think should be happening and it maybe it's overblown?
SPEAKER_01I think it's problems that are actually happening today because, you know, and again, there's not great. Meta's actually done some good work on this in root cause, but you know, the things that end up causing problems in optics, some of them are actually due to the fact that it's pluggable, right? Because you're plugging it in, there's the opportunity for dust to get in the connector and block the optical signal, right? So when you're swapping out a pluggable, if dust gets in there, all of a sudden you have a reliability issue, right? And it's attributed to the pluggable, but it's not the inside case of the pluggable that's the problem, really.
SPEAKER_01It's the fact that the pluggable, you know, there's something interfering between the pluggable and the and the connector. So when you go to CPO and you have a fully integrated solution, you know, that's done in a clean room facility, right? There's no dust there. Right. When you have things that are permanently attached or semi-permanently attached, you know, you're not dealing with these connection issues that you get where reliability concerns appear. So that's what the hyperscalers are really trying to become more comfortable with is not that CPO will be as reliable as pluggables.
SPEAKER_01That's not good enough. CPO has to be way more reliable than pluggables. So I don't think anybody questions that CPO is as good as pluggables. The question is, is it so much better than pluggables that it can be made a permanent part of the system?
SPEAKER_00Gotcha, sure. Fascinating. So it's almost it how surprising if the idea was like, oh yeah, pluggable is like the release valve where if there's a problem, you can unplug it and put something else in. But actually, what if the problem is the pluggable, the act of plugging it in itself? It's kind of funny. So, yes, if you can make something co-packaged in a clean room facility, maybe dust never gets in and you actually don't see those issues. That's really fascinating. So then um, what about like cost? I mean, everything has a trade-off.
SPEAKER_00So if if CPO potentially has great reliability and um it's much lower um from an energy cost, what about like dollars and cents? Is it is it more complicated? Is it expensive, or does it is it all a wash compared to like linear or or traditional pluggables?
SPEAKER_01Yeah, and again, it's it comes down to if you look at a pluggable, you strip away a lot of the components of the pluggable when you go to co-package optics, right? You obviously don't need the case anymore, right? That disappears immediately. There's a lot of support circuitry on there, um, some MCUs, some power, um, and in particular, especially the DSP that goes away, right? So the DSP is probably the biggest component that goes away from a cost standpoint. So that saves you a lot of cost.
SPEAKER_01Um, and then you get into the operating costs, and when you have significantly lower power, that obviously translates directly to operating costs.
SPEAKER_00Interesting. Okay, so um let's move on a little bit. So I saw some news about uh the OCI MSA, like a multi-source agreement for optical interconnects. So can you enlighten us? Like, what is an MSA? What is this? And why does the industry write one? I think the some of the names involved AMD, Broadcom, Nvidia, Meta, Microsoft, OpenAI. So like all ever it seems like everyone is a part of this. Um, why is this necessary?
SPEAKER_01Yeah. So an MSA is a multi-source agreement, and what it is effectively is the industry looking at a standard, typically a standard from the IEEE, and saying, okay, this is great, but there's a lot in here that maybe is options. And what we're gonna do is we're all gonna agree on here's the options we're gonna do, here's the options we're not gonna do, here's how we're gonna test things, here's how we're gonna make sure it's interoperable. You know, so that is is what an MSA is. The one that's it's not called an MSA, but honestly, the most famous MSA is the Wi-Fi Alliance.
SPEAKER_01So the Wi-Fi Alliance takes the Wi-Fi spec from IEEE and turns it into something that can be productized, right? The optical industry is typically at MSAs focused primarily around mechanical. So if you look at the, you know, there's one for uh it's called OFSP. That's the form factor for the pluggable, so it defines the mechanical electrical interfaces so that when you plug these things in, they all work, right? The electrical interface is defined by OIF and the optical interface is defined by IEEE, but the MSA kind of brings all those things together and says, okay, how's here's how we're going to combine these pieces and wrap them up mechanically such that we have an interoperable pluggable.
SPEAKER_01The OCI MSA is a little bit different in that it's actually taking and defining the optical interface, right? So that for most of the optical communications used in data centers, it's defined by IEEE, IEEE 803.3. That is not sufficient for what we want to do for scale up compute, right? And OCI is really focused on scale up compute. And so when you look at the participants, the founders of OCI, they are the key suppliers in the data center industry.
SPEAKER_01They know absolutely what is the best solution for scale-up networking, right? And so they've come together and said, hey, listen, the specs, the optical specs that the IEEE is defining are not really fit for purpose for the scale-up application. This is what we believe is fit for purpose for the scale-up application. This is how we think you should do things. And it's really focused around maybe not going as far range-wise, because you know, two kilometers maybe is a little too far for a scale up network.
SPEAKER_01So we're gonna back off on that a little bit, and we're gonna go to a wide slow solution. And what that is is you know, the to go as fast as possible with as much data as possible over a single fiber, IEEE has moved to multi level modulation. So it's called PAM4. And you know, you transmit two bits per symbol and you have four levels. And the issue is those levels are pretty close together, so you have to either have a lot of power or very low noise to receive them properly.
SPEAKER_01What they've done with OCI is they've gone back and said, we're gonna go back to a binary system, so only one bit per symbol, and we're gonna slow it down to 50 gigahertz. But we're gonna do four wavelengths per fiber. So we're gonna take advantage of the fact that you can put multiple channels on a single fiber to still have that 200 gigabits per second of bandwidth. We're just gonna do it in with four wavelengths. And what that does is it makes the receiver a lot less complex. So the amount of channel equalization you need goes down quite a bit, and the amount of forward error correction you need goes down by quite a bit.
SPEAKER_01The native bit error rate for NRZ 50 gigahertz is about a million times less than what it is for PAM 4. So it's significantly easier to receive an NRZ signal. And that translates into reduced cost and it translates into reduced power.
SPEAKER_00Oh, fascinating. Okay, so the industry is coming together saying we're gonna make an interoperable standard, and presumably that's just so that there can be competition so that everyone can say, hey, we're building to this spec collectively. Um and then what I heard you say, which is fascinating, is it sounds like the industry is aligning around, not just trying to continue to increase the per lane speed and double it, double it again, but actually to like walk back a little bit instead of pam for it, go back to NRZ and say, hey, let's go back to 50 um gigabits per second at uh NRZ, but can we send four um like wavelengths down the same fiber, kind of four colors, if you will?
SPEAKER_00And those are being sent simultaneously.
SPEAKER_01Correct.
SPEAKER_00Yes. Interesting. So if there's more, if there's different wavelengths being sent simultaneously, is that like a boon for like people that make the lasers? Do you need more lasers to do that?
SPEAKER_01So so just backing up one second, just to explain something about the difference between the scale up application and the scale out application. So the scale up application, what you're trying to do is you're trying to connect as many GPUs as you can using a single hop network. Right? And so if you look at the maximum size switch you get today is 100 terabits per second. And so with 200 gigabits per second per fiber, you can connect 512 GPUs to each switch, right? And then what you do is you stack up switches to give you the most bandwidth.
SPEAKER_01So if you look at the NBL 72 system, they have 18 switches in parallel, which allows them to get to that 7.2 terabits per second bandwidth they have per GPU. But each fiber is only, or in this case, they use copper, each copper pair is only 200 gigabits per second. It's just shared over 18 uh switches. The other thing that the OCI team has done really well is so when you look at the amount of lasers you need, it's not as simple as one laser per wavelength because the laser power can actually be split and used by multiple fibers.
SPEAKER_01Right? So if I have a transceiver that's got eight fibers connected to it, um, I typically right now the ratio is four to one, so I'd have two lasers to drive those eight fibers. But for pluggables, it's actually moving to eight to one, so you'd only have one laser driving all eight of those fibers. With the OCI spec, they've backed that off even further. You can get to as extreme as 32 fibers per laser. So one laser could drive 32 fibers. So you could get a full optical engine solution with the eight wavelengths of light that's needed for OCI with just eight lasers.
SPEAKER_00Gotcha. Okay. And then you were saying that we don't need to think about like one laser per port or anything like that. In fact, the laser, the light can drive um many different fibers, maybe is up to as many as 32. Correct. In like the most extreme case. That's pretty cool. Um, okay, so now tell about where does Global Foundries come in here? I know you guys announced scale and you called it the first OCI capable platform. So tying back to that MSA, can you explain a little bit more about scale and you know how you're the first OCI capable platform?
SPEAKER_01Sure. So what we're doing with scale is we're putting together everything you need to do the electrical to optical translation in a known good form factor, right? And call it a triplet, call it a module, whatever, right? And so what that includes is it includes an electronic IC, which is going to be used to communicate to the CPU or the GPU or the AI ASIC. And then that electronic IC sits on top of a photonic IC, which does the translation between electronic and photonic signals.
SPEAKER_01That photonic IC has a detachable connector on there, and that detachable connector is key because when you're assembling a CPU, what you don't want to have to do is have a fiber dongle hanging off your photonic IC at all times. You want to create, you know, you want to build your server using rectangles, right? And then plug everything in at the end. And so with that detachable fiber connector, it looks very much like copper, where you build all your PCBs, put them all together, and then you have a wiring harness, in this case an optical wiring harness instead of electrical wiring harness that you connect at the end that bridges from those CPO modules connected to the GPU to the edge of the server.
SPEAKER_00And global foundries, you you can fabricate. Which of those pieces do you guys fabricate and assemble?
SPEAKER_01Yep. So we have the ability to manufacture all of it. Um, the photonic IC for sure, we do 100% of the manufacturing. Um, we have a number of different design partners working with us on photonic IC designs. The electronic IC, it really depends on the complexity of the design, particularly the digital. So if it's a less complex design, we can use our FinFet process or one of our FDX processes. If it's a simple, you know, linear translation, we could actually use our Silicon Germanium process even.
SPEAKER_01Um if it's a more complex digital translation and a customer wants to use an advanced node, like a three nanometer or two nanometer node, then they're free to do that and we'll bring that wafer in and do the assembly and test. The micro optics um again are manufactured externally, but we do the assembly and test on that. And so what we deliver at the end of the day is a known good, fully tested module with very high reliability.
SPEAKER_00Hmm. Okay, fascinating. So then talk to us more about like position yourself against what some other people are doing. Like I know TSMC is in this space with Coop and they use a grading coupler. And I think um Global Foundries uses a different technology, an edge coupler. And could you maybe explain a little bit more about the technology differentiation that and the approach you're taking?
SPEAKER_01Right. So we do so, yes, the mechanism for the light getting into the waveguides is very similar to edge coupling, but we use we don't do it at the edge of the die. We actually have a micro mirror that we implant into the wafer so the light can still come in vertically, right? So the light comes in vertically exactly the same as you would with a grading coupler. We just have a mirror and that reflects the light horizontally into the waveguides on the photonic IC. The reason we went with that approach instead of grading couplers is because grading couplers inherently have limited bandwidth.
SPEAKER_01Um, and so you can't service the entire O-band with a single solution. With a micro mirror, we can do the entire O-band, which gives us the maximum flexibility as far as doing the wavelength planning. Right? So OCI today is four wavelengths in each direction. You know, one of the ways they're going to scale that to do higher bandwidth is to add more wavelengths. And so as we're thinking about how do we evolve into the future, we wanted to make sure that we made the entire O-band available to use for as many wavelengths as possible.
SPEAKER_00Ah, fascinating. Oh, and for listeners who might not know, can you define like what does O-band mean?
SPEAKER_01Uh yeah. So the O-band is it, you know, you think about AMFM bands for your radio. So O-band is an optical band. It's around 13, 10 nanometers, um, which I can't divide by the speed of light in my head to figure out how many gigahertz that is, but or terahertz is, but it's really, really fast. Gotcha.
SPEAKER_00Okay, thanks. That's helpful. All right. So then, okay, so you guys are taking a different approach, and you think, you know, it keeps the whole O-band on the table, which will be helpful in the future as you're trying to increase bandwidth, it will keep more wavelengths sort of at your disposal. Um, at the end of the day, you know, I see people online and they'll say, you know, oh, this one company is better than the other, um, but it feels like a little bit sort of like armchair quarterbacking. Like, like, can you help us understand as people interested in the space, like, how do we know as a non-engineer, uh, you know, non-silicon photonics engineer, how do we know how good you know, Global Foundry's components are?
SPEAKER_00Is there like a good proxy or a good way to understand? Is it just like looking at the customers, or what do you think?
SPEAKER_01Yeah, I mean, that's the you know, the market doesn't lie, right? So the market will choose the best solution, right? And so that's you know, we're at the very beginning in this market. Um, and like I say, right now, you know, TSMC, other foundries, they're not the enemy to me, right? The enemy to me right now is copper. I'm trying to be copper, right? If TSMC wins and we win, that's great because we're both displacing copper. And until all the copper is gone, there's plenty of market to go around. So sure. Um, but from a standpoint, yeah, how do you how you know how are you gonna tell from the outside?
SPEAKER_01You know, you will you will see customers adopt certain solutions, and and that's gonna tell you who's who's got the best solution.
SPEAKER_00Gotcha. That's helpful. Um, so then also one other thing, I think I saw on a press release somewhere um about scale mention of known good dye testability. Can can you explain that a little bit more for listeners?
SPEAKER_01Yeah, and known good dye is probably under calling it, and I mentioned earlier, it's known good, very, very high reliability dye. So with known good dye, what we're doing is making sure that what we deliver to our customers is 100% functional, right? And that's a very complex process because we're looking at a solution that has electrical signals on the bottom, so you can connect it to the substrate or the interposer, and the optical signal coming in from the top. So normal electronic test is only one-sided.
SPEAKER_01You you you know, you drop your pro pads on and you test things and you pull it off. We've got to do a two-sided test, right? We've got to take light in from the top, we've got to put electrical signals in from the bottom. We have to test that at high speed, because we're talking about signals at 200 gigahertz coming in, and we've got to test a lot of channels because these optical engines, you know, a normal uh pluggable transceiver right now maxes out at 1.6 terabits per second. These optical engines are gonna be 7.2, 14.4, even higher, right? So we're talking about a lot of bandwidth that has to be done in parallel.
SPEAKER_01So we've put a huge effort into you know building test systems that A can connect top and bottom, and B have enough bandwidth to do this full testing. And so we'll deliver that known good optical engine to the GPU provider or the ASIC provider, and they'll integrate that onto their package, very very similar to how they would do an HBM today or an IODI today, right? That's gotcha, and then and that testing gets you to you know time zero reliability, right? Time zero yield.
SPEAKER_01But now you have to have reliability, right? You need this to survive in a very, very extreme environment for a very long time, right? You can't fail in the field, you can have zero failures, right? So beyond just the kind of time zero testing and make sure it works, you know, the first time you plug it in, we need to make sure that it works for 10 years plus under very extreme circumstances, right? And that reliability testing is actually key to success long term.
SPEAKER_00Nice, that makes sense. Interesting. So when I'm zooming out and I'm thinking about you know the technology you have and the ability to offer the pick, the photonic IC and the um electronic IC, or package that in from someone else and do all the coupling. Um, and then of course we're talking about scale up, where this feels like it's going to be a very large TAM because um just there's you're have the opportunity to connect lots and lots of GPUs. So there could be lots of uh optical engines, lots of fibers, um, big opportunity there.
SPEAKER_00Like, what are you guys saying publicly about how big you think this market is and and how much opportunity there is for global foundries with CPO?
SPEAKER_01Yeah, I mean if you look at the scale up bandwidth compared to scale out, right? The scale up bandwidth is somewhere between five and ten times what scale out is, right? And the scale up today is all copper. So if we can convince everybody to move their scale up networks from copper to optical, you know, we're sort of instantly talking about increasing the size of the market by 5x or 10x. Right? That that's why I say the enemy is copper, it's not other foundries, right?
SPEAKER_01We want we want to enable that conversion, we want to work with other people, we want to partner everything we can to convince the market to move from copper to optical for scale up, you know, and and because it's the right solution, right? So that's that opens up a huge potential market. So if you look at, I can't remember what the latest projections were from from light counting or signal AI, but you know, tens are close to 100 billion in optical, you know, think of that market going 5x, 10x bandwidth. Now, it's not going to make the market go 5x or 10x because, like I said, the the CPO solutions have to be much more cost effective than what you're looking at for pluggables.
SPEAKER_01But you know, doubling easily 4x, maybe, and then pile on top of that, you know, the bandwidth continues to increase, you know, the number of servers continues to increase. So, you know, it's it's a very high growth market for a for a while.
SPEAKER_00That's promising. That's exciting. Um, okay, one other topic. You mentioned silicon germanium in passing. And I I think I saw on your most not today's earning's call, but the most recent one before that, a quarter ago, um, mention of silicon germanium capacity in Vermont being oversubscribed. And anytime I hear something oversubscribed, you know, my ears kind of perk up. So can you tell us like what is the silicon germanium used for? And like, yeah, it enlighten us a little bit. And why is it in such demand?
SPEAKER_01Right. So when we're looking at optical transceivers today, there's two components that are used to locally on the pluggable or even between the pluggable and the GPU to amplify the signal coming into or going out of the photonics, right? One is the driver, which is a it's a modulator driver, so it it amplifies the signal coming out of the DSP such that you can drive the modulator that's the silicon photonics modulator on the pick.
SPEAKER_01That to some degree is being integrated into the into the DSP, so the market's shrinking a little bit on that. The other one is a transiped amplifier. So that takes the current out of the photodiode, which detects the incoming light and translates it into a voltage, which is then converted and processed on the DSP. The bandwidth of those is similar to what I talked about earlier with the PCB. You know, for a 200 gigabit per second PAM force signal, you're talking about 55 gigahertz of bandwidth if you've got a re-time system and closer to 70 gigahertz of bandwidth if you're trying to do a linear system.
SPEAKER_01Typically, for an amplifier, you need about the transistor speed needs to be 5x what your bandwidth is. Right? So now you're talking about transistors that need to run 350, 400 gigahertz to do the 200 gig per lane. When we talk about 400 gig per lane, now you're talking about doubling that, right? So you're talking about transistors that are 600, 700 gigahertz. Very, very few companies can do that at scale and reliably. And silicon germanium from GF is one of those few technologies, right?
SPEAKER_01So we have you know very, very good silicon germanium technology. Our transistor speeds are extremely high, much higher, you know, higher actually than what you need for 200 gig per lane. And we have new generations, you know, roadmap to support 400 gig per lane very quickly. So, you know, we've been fortunate that we've gotten a lot of adoption on 400 gig per lane. So as you see the 1.6 T transceivers volume increasing, you know, we're getting a disproportionate that share of that right now.
SPEAKER_00Okay, interesting. So silicon germanium is about transistor switching speed. And as we see switching speeds uh or bandwidths of like 200 gig, 400 gig, it means we should think silicon germanium and we should think global foundries.
SPEAKER_01Yes.
SPEAKER_00Fascinating. Awesome. All right, Tom, we've covered a ton of ground. It was pretty technical, but I appreciate you walking us through and teaching us and enlightening us also about global foundries from a business perspective. So thanks so much for your time.
SPEAKER_01Thank you, Austin. It was great talking to you. I appreciate the time.
Austin Lyons and Vik Sekar break down Google's two new TPU systems, the 8i Max and 8TX. These are not new silicon but new system configurations of the existing V8 chip, designed for specific workloads. The conversation...
Jensen Huang resigned as Nvidia CEO this morning, and Austin and Vik recorded two hours after the news broke. They argue that a board must prioritize corporate integrity over short-term shareholder value, even for a...
Intel just had its worst day on the stock market in 55 years, with shares falling nearly 30%. But here's the strange part: Intel actually reported a strong quarter. Revenue grew, its data center business surged, and the...
The fundamental physics of data movement are forcing a shift to new memory architectures for AI inference. Sudeep Bhoja, co-founder and CTO of d-Matrix, joins Vik to detail the company's journey from a non-consensus bet...
Marvell's potential acquisition of Lumentum is a strategic vertical integration play to capture the massive value optical components now unlock in the AI era. Andrew Schmitt and Scott Wilkinson of market research firm...