GlobalFoundries Thomas Barber: CPO, Silicon Photonics, 300mm, SiGe, OCI, NRZ

Good morning. How are you?

Sure. Yeah. And I think, you know, GlobalFoundries' 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 all the things that are essential to making these electronic systems work. And silicon photonics perfectly aligns with that strategy. You know, it's very differentiated. It's very complex technically. And it's in a very challenging market, right? So there's a great opportunity to differentiate. For those people that don't know, you know, GlobalFoundries, we were born out of AMD. 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 IBM Microelectronics actually started in the mid-2000s with, you know, doing very basic integration of waveguides, modulators and photodiodes into CMOS technologies in 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. 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.

Yeah, 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, 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 photonics to 300-millimeter foundries, right? We put it into our facility in Malta, New York. 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 everybody's just desperate for more capacity, and we are available to provide that capacity. As 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.

Yeah. So, yes, most silicon photonics today is actually still produced on 200 mm wafers. That's where most of our competitors are. When you go to a 300-millimeter wafer, you know, it's the radius squared, right? So you go up by 50% in radius, you square that, you know, you get 2.25 times as many die 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? For 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.

Yeah. So it all comes down to the fundamental limitation of copper, which is range, right? So at 100 Gbits per second, a direct-attached copper cable, so no intelligence, just a dumb cable, could span about 2 m, 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 Gbits per second, that range drops to about 1 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. When you go to 400 Gbits per second, now the range is half a meter, and you're not going to quarter the size of the rack, right? So you have to go to something faster or 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 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. You cannot do that with copper. Not at the speeds we want to do.

Yeah. 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 count them, you know, they're doing 300, 500, 1,000 GPUs across, you know, 6, 8, 10 racks, right? So they're really expanding the size of their scale-up networks using optical communication to span multiple racks.

Right. So it can be any of those options. What we see is a trend from pluggables first going toward NPO and then towards 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 DSP in the pluggable. When you move to an NPO system, now you've dropped your loss from about 35 dB to somewhere between 15 and 20 dB, right? That allows you to significantly reduce the amount of compensation you need to do on the electrical side, even eliminate it, right? So if you have enough link budget end to end that you can handle the extra, you know, 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 retimer 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 saves a lot of power. But you still have a fairly bulky package for this NPO, and 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 need 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 SerDes significantly. Right? So 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 in 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 SerDes to overcome that 15 dB of loss between the core processor and the NPO. When you go onto the package with the GPU or the ASIC, now you're talking about 3 dB of loss, and now you can go to the lowest power SerDes, which allows you to get down to less than five picojoules 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 for 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.

Yeah, what it is, it's the electrical traces through the PCBs from the core processor, you know, couple hundred centimeters to the edge of the server, right? And so those are very thin wires. And when you're talking about, you know, a 200 gigabit per second PAM4 signal has about 55 GHz 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.

Right. 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. But if you swap out the pluggable and the problem goes away, then you're fine. You keep going. Moving to near-package 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 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. So 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 has 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-packaged 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. So, you know, 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.

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 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. It's the fact that the pluggable, you know, there's something interfering between the pluggable 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. That'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?

Yeah. And again, 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. Some MCUs, some power, 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. And then you get into the operating costs. And when you have significantly lower power, that obviously translates directly to operating costs.

Yeah. 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 going to do is we're all going to agree on here's the options we're going to do, here's the options we're not going to do, here's how we're going to test things, here's how we're going to make sure it's interoperable. You know, so that is what an MSA is. It's not called an MSA, but honestly, the most famous MSA is the Wi-Fi Alliance. So 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, there's one for, it's called OSFP. 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, here's how we're going to combine these pieces and wrap them up mechanically such that we have an interoperable pluggable. The OCI MSA is a little bit different in that it's actually taking and defining the optical interface, right? So that for most, the optical communications used in data centers is defined by either 802.3, 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 you look at the participants, the founders of OCI, they are the key suppliers in the data center industry. They 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, 2 km maybe is a little too far for a scale-up network. So we're going to back off on that a little bit and we're going to go to a wide slow solution. And what that is is, you know, to go as fast as possible with as much data as possible over a single fiber, the industry 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. What they've done with OCI is they've gone back and said we're going to go back to a binary system. So only one bit per symbol and we're going to slow it down to 50 GHz, but we're going to do four wavelengths per fiber. So we're going to take advantage of the fact that you can put multiple channels on a single fiber to still have that 200 Gbits per second of bandwidth. We're just going to do it 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. The native bit error rate for NRZ 50 GHz is about a million times less than what it is for PAM4. So it's significantly easier to receive an NRZ signal. And that translates into reduced cost and it translates into reduced power.