Hello everyone. Today we have a special guest, Al Yuen, CEO of Picogiga. Picogiga is an optical connectivity company and we'll get into all the interesting details. But first I wanted to introduce you guys to Al. So Al, tell us about you and your background. I know you've been in the industry for a long time.
Yes. So after grad school at UC Santa Barbara, where a lot of the photonics folks have originated, I went into HP, HP Labs, where we focused on photonics research, etc. And since then, around '99, I left HP and started my first company called Alvesta and we created the world's first 10 gigabit Ethernet. At the time, 10 gigabit, which is 10 billion bits per second, we were actually trying to figure out applications that how people would use this in '99. We would make up things like people wanted to stream video someday in all the rooms in their house. And so we had to actually... but today, of course, we're doing 1600 gigabit or 1.6 terabits. So since then, I've gone to various companies. I ran a division for Coherent. And then started some other solar companies and clean tech and then finally ended up at Lumentum in my last gig. And then about 2 years ago, through Playground Global, which is our funder, we started Picogiga and so far so good. And what we're doing right now is basically in the interconnect space and we'll talk more about that today.
Awesome. Wow. What a great background. I love that you guys invented early 10 gigabit Ethernet and then had to create ideas to sell people, to convince people that like, yes, this is useful. Yeah, people will want to use it. That's awesome. Now, remind me, you also helped invent the active optical cable. Is that right?
Sure, yeah. So, very interesting story. So, back in Alvesta, my first startup, again, another small startup at the time called Mellanox, which of course now is inside Nvidia and created this whole kind of hyperscale and, you know, the whole InfiniBand, etc. So, they approached us. So, they had these very bulky copper cables, even back then, right? So, 20, 25 years ago. And they said, "You know what? The copper cables are very bulky. They could only reach, you know, at that time, tens of meters. Now, it's even shorter. But, they said, 'We would like an optical option, but we don't really want to commit to a full kind of optical solution. So, could you put the optics inside the connector?'" And we said, "Why not?" right? So, basically, we took that same transceiver that's typically on a board, and we literally embedded it right into the connectors, you see here. And then we thought, "Well, that's kind of, you know, not very efficient and clean." And so, we embedded that whole connector inside. So, the optics went inside the connector, and then this is the world's first demo of an active optical cable, meaning electrical to electrical. So, electrical comes in, electrical goes out, but inside the electrical to optical transition. So, electrons come in, photons carry the information, and then electrons go back to the point B. So, that's how the whole active optical cable concept came through Mellanox. And then, since then, for the last 25 years, the AOC has been the standard workhorse in many, many data centers today.
Amazing. So, having invented that and then watched it sort of become just widely adopted and probably produced, you know, I don't know, in the millions of cables or something. How has that impacted the way you think about like what's possible as an entrepreneur and in this space?
I like to think of ourselves more as engineers, right? So, there's a difference between, I would say, scientists or researchers or what people call R&D, right? Research and development and engineering or product development. And so, I always kind of tell people I'm more of an engineer. And engineers solve problems and they want to basically create products that basically meet specifications, right? So, if you need a Prius, you certainly don't design a Ferrari, for example, right? So, that's a little bit overkill. And so, therefore, the product fits the need, the spec, cost, reliability, in our case, reach or the amount of power, etc. So, in answering your question, I feel like what we do is really look at trying to solve the specific problem. And today, right now, copper has shrunk to about 3 m, 4 m reach at 200 gigabits per second per lane. And therefore, they can't get the information off each rack, right? So, racks and racks of these GPUs, CPUs, lots of compute power. So, these racks are getting very hot because they're packing more and more GPUs per rack because they can't exit the rack. And so, to exit the rack, you need an optical solution. And so, there's many technologies and of course, VCSEL-based active optical cable is one of them and that's what we're basically focused on, solving this problem of low-cost, highly reliable, again, going way back to our roots, which is replacing a copper cable with an active optical cable. So, the problem really hasn't changed. It's just that the speed and the aggregate bandwidth is now 1,600 times more than the old 1 gigabit Ethernet. So, it's exciting. It's been a long journey, but we're still at it and we see a long future for VCSEL-based technology as we go forward.
Nice. Nice. Okay, so you have this history as thinking like an engineer of just like what is the problem that's right ahead of us in the industry and not saying like we need to invent new physics, but just like how can we be thoughtful about, you know, oh, same cable form factor for example in the AOC case, electrons in, electrons out, but we could use optics here and could we put the transceiver, you know, on the end of the cable. So, you know, thinking very pragmatically. And now, you know, fast forward 25 years or so and I guess in 2024 you started Picogiga and it sounds like you're going to tackle the problem again of communicating lots of information, this time it's, you know, GPUs that are talking to each other like rack to rack where it's such high bandwidth copper is shrinking to 3 m or so and so you guys again are thinking about how do we use essentially active optical cables, but this time with VCSELs. So, tell us like why VCSELs and what about the problem made you want to start Picogiga to actually start a company and like get in the game here?
Great question. So, VCSELs for one thing has been around since '96, right? So, it's a technology that's been in product in the field in data centers since 1996 starting with the first gigabit Ethernet. So, you know, I'm more of a historian today. So, here's a 1 gigabit Ethernet that, you know, HP created and at the time Honeywell etc. Very very early companies that did 1 gigabit. And so, from the 1 gigabit all the way to today, again 1,600 gigabit or what they call 1.6 terabits has all been VCSEL-based. And going back to what I was saying, the practicality of these solutions has to meet capacity, demand, cost, all of those, right? You need to check all those lists. If you're trying to get into a hyperscaler today, you need to basically meet all that checklist. You can't say, "Oh, I meet everything, but, you know, it's three times the cost of your per target." Or everything is great, but, you know, I can't get the reach. So, they want all of that met. And so, right now, the only thing is copper, right? And that's mainly driven because of the cost. Cost for copper, of course, is very, very minimal. Even active copper, where you have some signal integrity, signal processing, etc., built in, like an AEC, active electrical copper, the cost is still quite minimal compared to other technologies that are more elegant, longer reach, etc. So, getting back to your question, what's very important in a data center, if you have different fiber optic communication, I think everybody hears it and goes, "Oh, isn't that the over the ocean, they have these transatlantic subsea type of cables?" And said, "Absolutely." And so, there's many different flavors of optical communication. So, when you say that, it doesn't cover all of them, right? So, it's a very big umbrella, and what we're talking about today for data centers is a very short reach. Today, they call it scale up. So, it's basically a row, typically 25, 30 m, right? Or in English terms, 75 ft, okay? So, it's a very short, you know, within your house, basically from one end to the other end, for potentially even shorter. So, in that very short reach, what happens in communication is the shorter the reach, typically the higher the volume, right? So, if you think of it, you have a lot in the data center where you're trying to connect all of these different GPUs, CPUs, or ASICs together. And so, there's millions of interconnections, miles of fiber inside a data center, which may be, you know, a football field length today. But, when you leave that data center, and you go off to between cities, between buildings, between countries, you have fewer fiber, but you need longer distance. So, getting back to the data center now, and long kind of answer to your question is, you have to have millions per month capacity to meet the volume demand. So, that's one of the boxes, right? You got to check off all the boxes. So, if you're able to demonstrate one or two racks, or go to a show and demonstrate this technology, again, technology, science, R&D shows the capability. Let's call it a demonstration. But, in order to ship in volume, millions per month, you need this whole ecosystem. You need people who make the connector, you need people who make the transceiver, you need people who make the sockets, right? So, everything has to be millions per month, and any one of those bill of material parts that go into these transceivers, if it's missing, then you can't really ship in millions per month. And that's what's happening with a lot of these new technologies that are fantastic, by the way, right? From a demonstration, from a capability going into the future. But, VCSEL's domination is that the last 25 years we've been shipping in the millions per month. So, therefore, there is no invention of a technology or capacity or foundries in order to build these up. We already have that. So, now we design the latest VCSEL. So, today it's 200 gigabit, which we just announced. And then we put it into the whole ecosystem, and they package together, and then voila, we can build millions per month without waiting for machines to be, or actually even buildings to be built up, and then new machines, and then new process to develop. All that has existed for the last 25 years to service this very short reach data center application.
Mhm. Mhm. Okay. So, yeah, let me reflect it back to you. So, the problem that you're trying to solve is scale up with optical interconnects. And but the key insight that you have kind of with your pragmatic hat on is how can we use technology and a supply chain that already exists and can already ship millions of cables, components, parts per month. And so, that's one reason why VCSELs is so attractive is because they're not a new technology. The supply chain is not new. So, even though as we hear about all these other interesting things, because I think anyone paying attention, you know, listens and they say, "Oh, you know, the Broadcoms or Coherents or Lumentums." I hear about silicon photonics. I hear about EMLs. You know, I hear people talk about micro LEDs. But you're saying like, "Hey, don't rule out VCSELs even though they're not new, they're still there's advantages to VCSELs." But tell us like, if VCSELs have been around for so long, like why aren't other folks trying to take VCSELs to the 200 gig per lane * 8 lanes 1.6T and beyond?
Yeah, so like you said, there's multiple technologies, silicon photonics, EML, even this micro LEDs right that are coming. All vying for this 200 gigabit type of electrical signal coming in. So, from the ASIC, GPU, CPU, you have 200 G per lane of electrical signal, right, that's coming in. Now, when you go to the electrical to photonic or E to O type of transition, then it can go to different lanes through, again, a different IC that may be a gearbox, they call it. And so, you don't necessarily have to run at 200 G straight through, which we can, right? So, that's the most elegant because if you can go straight through without having to go through an intermediary gearbox, then it saves power, cost, and then latency, which is the delay going through there. So, 200 gigabit would be ideal. But, what's changed in hyperscale is different from Ethernet. So, Ethernet information is sent in what we call packets. So, whatever information, oh, where should I go in Italy? I'm going on vacation, da da da. So, that information is divvied up by your search engine and sent in packets. And then that comes back to you. And we're not very cognizant if there's an error drop or there is some delay because those are in the, you know, hundreds of nanoseconds or milliseconds. And to us, it's like, you know, 1/1000 of a second. We're just not going to delay. But, for hyperscale systems, which are literally made up of thousands of GPUs acting as one brain, if you will, right? One supercomputer, one high-performance kind of cluster of computers, then that latency is super critical because the GPU notices anything in the tens of nanoseconds or hundreds of nanoseconds. So, therefore, the typical Ethernet bit error rate, which is the measurement of how much errors you're getting per second. And so, if that needs to drop from 10 to the minus six to below 10 to the minus 10 or what we call error-free today because any errors then slows down the whole training inference all the AI infrastructure that's needed. So, that has changed and allowed these very, very, you know, typically higher cost single mode, oh, I'll mention that. Single mode being longer distance, longer reach, very high performance for long distance, now has come into the data center and to the hyperscale systems because of this requirement for very low bit error rate, very high performance computing and connectivity. And now VCSELs has to raise its bar from 10 to the minus six typical Ethernet the last 25 years to 10 to the minus 10, 10 to the minus 12. And we've done that, right? So now we've pushed our VCSELs to 200 gigabit even though you can use it at 100 gigabit and 50 gigabit NRZ, you can still leverage that to very low bit error rate. So the answer is that things have changed for hyperscalers to require error-free and that's allowed all these very high-end single-mode solutions now to compete directly with VCSELs because of that additional specification that's new to hyperscale AI systems.
Okay, okay. So because the capacity for sustaining errors is much much lower, we don't want all these GPUs to just be waiting. That sort of changes the game from like the cloud SaaS day to now when everything's acting as one big computer. And so this bit error rate has to be much much lower. And so VCSELs, which are short reach, have to either essentially come down to meet that or what you're saying is these other technologies that are longer reach, higher power, those already were closer in the demanded or like the necessary bit error rate. And so people are saying, "Oh, why don't we take those technologies and bring them into short reach?" But surely that has like power and cost tradeoffs of taking something that could talk at low error rate like over a long distance and trying to bring it in. And so you guys must be taking a different tact and saying like, "No, no, no. Like let's just make VCSELs error-free essentially." And presumably that's like a cost or a power tradeoff that you'd rather go with? Or is it back to the manufacturing supply chain capacity?
Yeah, so it's again, it's a quite complicated matrix of items you have to check off, right? So, those who are shipping silicon photonics and EML for long reach or DFBs, etc. Single mode, very high performance devices have been around for again, similar time, 25 years and they've been used for long reach because they have that super high performance, very few fibers and then what we would call WDM, wavelength division multiplexing. So, fibers are very expensive when you're going hundreds of miles, hundreds of kilometers and so you want to use very few fibers, but then pass more information through more wavelengths, more colors in that same fiber. For again, short reach, when you're talking about tens of meters, you're not as kind of locked into the cost of the fiber, it comes down because you're only 10 m versus 10, you know, kilometers or extending. So, therefore, the capacity for volume supply chain of the very, very high performance, low bit rate would be a natural, right? They come in, hyperscalers want low bit rate. Let's go with the Ferrari, right? Let's go with the super high speed, you know, high performance single mode, but they've been used to building in the maybe 100K, 100,000 kind of volumes because you don't need as many of those between cities, between countries. And all of a sudden, they come into the data center, even though their performance is excellent, cost is a little bit higher because of single mode nature of packaging. However, just the infrastructure to build millions per month now is 10, 20, 50X what exists. So, that's brick and mortar. That's basically saying, I only have one pizza oven and I've been used to, you know, a small clientele that I can make maybe about 20 an hour. Someone goes in, I'd like to place an order for 5,000 pizzas and I need them in an hour. So, you're going I need basically 100 pizza ovens. That's the exact same problem that the very high performance single mode long reach traditional having coming into so silicon photonics EML excellent excellent technology very very good bit error rate all of those but then the infrastructure needs to be built up and that's what you're seeing, right? You're seeing a lot of announcements with people holding shovels and saying we're investing in the next, you know, supply chain the buildings etc. and that's great, right? Certainly bringing more manufacturing not only in the world but back to the US and those are all great for the industry as a whole but VCSELs have been around for 25 years and shipping in the millions. So, we're not building we're not putting shovel to ground etc. We're just changing the actual VCSEL performance the chip and then leveraging the existing infrastructure and so what we call for VCSEL capacity it's unconstrained, right? So, there's constrained meaning they're sold out we're sold out through next year. If you're going to place an order it's going to be after, you know, be a eight month 18 months lead time a year and a half from now we can get it to you. For us unconstrained just means we have a certain lead time that's basically only limited by our cycle time of building through the factory. So, if for us, you know, a VCSEL run and then a packaging run maybe four eight 12 weeks but that's limited just to the fact that we have to build it out and ship it but it's not limited by the constraint of the supply and ecosystem of machines or pizza ovens. We have plenty of pizza ovens. Place the order we'll get you your order in the cycle time that we commit to.
Gotcha. Okay, that's very interesting and a great sort of competitive advantage there for you. So, on the constrained side, is that what we hear about for listeners like when they hear about like indium phosphide being a bottleneck and just like or is it partic- Like where in the supply chain is it constrained? And then for you, like what's different about VCSELs that does make it unconstrained?
Yeah, so for indium phosphide, a lot of the silicon photonics and EMLs, etc. is based on this material, indium phosphide. For VCSELs, our technology has always been gallium arsenide. I know for the listeners that may be okay, one three five compound versus the other, what's the difference? Indium phosphide is material constrained from the very beginning. Like you can't even get a base substrate, right? Even before you process it, just the substrates for indium phosphide are limited before you get it all done and basically made into a product, whether it's EML, silicon photonics or VCSELs. The indium phosphide substrate, bare material is already limited. Gallium arsenide, unconstrained. So, we start with that. And now you go through the fabs and then of course fabrication, etc. Depends on foundries. They're usually very large companies that do foundries, companies like Picogiga and others. We don't own large clean room factories that make these. We design the VCSEL. We design the individual kind of VCSEL chip device. And then we use foundries to manufacture, right? So, those foundries are available, but they can't get enough indium phosphide starting material to do that. Now, after that, again, once you get to the chip level, right? You dice up, you go, "Okay, great. I've got the laser. I'm ready to go." Now, to get from the chip to a pluggable device, right? To an actual optical engine, if you will, there's a ton of stuff happens, right? Now, you've got laser drivers, you got boards. And then for single mode, you have to have all the machines that align that particular silicon photonics EML to a very, very small core single mode fiber. So, those machines have to be readily available, etc. And so, VCSELs again has that millions per month type of volume infrastructure, doesn't need to be built up. And then we get back to from the very beginning indium phosphide material constraint. And then you have to build it into lasers and then finally have to package it into transceivers. And at all along that supply chain, it's not used to building millions per month. So, all of that has to be built up like hardware, machines, alignment machines, testers, etc. to get there. Again, VCSELs has all of that infrastructure existing already.
I see. Yeah, that makes a ton of sense. I love your props, by the way. I like the little VCSEL that you held up. So, tell us more. So, you Yeah, that's awesome. Maybe you tell us what we're looking at and then walk us through exactly what you design and then kind of where it gets handed off and built and packaged and yeah, where your responsibilities end.
Yeah. So, basically, again, background is super important, right? So, there we call kind of this tree of knowledge of VCSELs, right? So, you have this line from Honeywell through Finisar and then Finisar goes into 26 and then goes into Coherent. So, there's a Coherent line. And then for us was obviously HP went into Avago, went
into Broadcom, and there's this kind of HP Broadcom line. And then finally, there's this, I would say, Picolight E2O. Again, throwing in some old names from 25, 30 years ago. They go into JDSU, which is another big name in the 2000.com time. And then JDSU spins off Lumentum and Viavi. And so, we're in the Lumentum arm. So, all of these three major arms have a lot of VCSEL knowledge. And so the strength of Picolight is we've tapped into and we have people, designers from all three of these branches. So imagine that all the know-how, again, not patented know-how about recipes and, you know, I use the example you can just hand three different chefs a recipe for making souffle and most likely you're going to get three different types because it's really difficult to get perfect souffle, if you will. It's not just getting some, oh, you get the eggs, you crack the eggs, you beat the eggs, right? So it's a lot more to that. And the same thing is in the VCSEL. So answer your question, what do we do? So we take all that know-how and then we design the epi layer. So you see all these little kind of lines. So these are epi layers that are designing this vertical cavity. So it's not an edge emitting. So edge emitting basically is a flat chip and then the light comes out of the edge. Surface emitting, right? VCSELs, vertical cavity, surface emitting, comes out of the surface. And so we design all of the kind of an internal cavity of the laser, all the dopings, and I'm not getting into too much detail, but all the process. So after we design, we hand it over to an epi foundry that grows this material. They give us back an epi wafer that's unprocessed and then we've been working with Win Semiconductor in Taiwan. That's our foundry. So once that epi wafer's ready, we hand it over to Win and then they process it, they process it into their clean room process and then they make the actual final VCSEL device. And then the beauty of the VCSELs also compared to an edge emitter is at the wafer level you can start testing and probing each one of these, 100% tested, what they call known good die, before you have to singulate it and dice it up into arrays, etc. So, that advantage is huge because if you have to add, the more work you add in before you quote yield the device, whether determine whether it's good or not, the better the cost, right? Cuz you always want to yield upstream. The more value you add and then you yield downstream, you lose all of that value added. So, wafer level testing for VCSELs is really, really advantageous versus the other kind of edge-emitting type of technologies.
Nice. Man, I'm learning so much history and so much sort of VCSEL 101 here. That's awesome. So, you work with Win, you design the VCSEL, you have an epi partner who helps with that. Win is able to even do the testing at the wafer level because it's a VCSEL, known good die. Ultimately, it gets built into the device. So, they must be doing the packaging, like some level of packaging for you, too, Win?
Right. So, Win is only on the wafer processing. So, they come to the wafer and then it's diced up into individual VCSELs. And these individual VCSELs, we work with our partners to build into either active optical cables, right, or transceivers, etc. And there again, another foundry, if you will, but this is a packaging company. However, with that said, companies like TSMC, right, Taiwan Semiconductor, now is going into co-packaging, meaning after they make their silicon wafer, they'll start packaging the optics directly on top of their wafer, which is co-packaged. So, there's another whole field that's growing called CPO, etc., where the traditional semiconductor foundries that are kind of wafer processing, now are stacking up different technologies together and they call it 3D kind of wafer level kind of packaging. So that's all emerging. For us, we only do the wafer at Win and then that Win VCSEL goes to our module integrator partners and then they'll build it up into the active optical cables or transceivers.
Got you. That's helpful. So then take us back to your road map. I know you mentioned a 50G version, 100G, 200G. Can you tell us more about... And I know you also mentioned a recent launch. So tell us more about your road map, what you launched, what you announced.
As I mentioned early on, 200 gigabit per lane is kind of the benchmark, right? It's the bar you have to clear. And EMLs, silicon photonics have all done that. And now VCSELs have reached that. We will just announce our 200 gigabit will start sampling next quarter. And so that's for a very simple transceiver where you have eight channels of 200 gigabit coming in. The aggregate or the combined bandwidth of that eight channels of eight by 200 is 1,600 gigabits or 1.6 terabits. So that's the standard ramping today, right? There's 800 gigabit transceivers as well. That's also shipping. That's eight by 100. And then the next generation or today's generation is eight by 200. So that's the 200 gigabit VCSEL that we announced. However, there's many, many different flavors of that because of the specification. So aggregate bandwidth 1.6T, right? Check. But there's different ways to get there if you want very low power or very low bit error rate. So running 200G, I liken it to a Ferrari, right? It can give you the 200 miles per hour, but it's a very high-end, relatively expensive because you need certain signal integrity, signal processing, all of that that adds to that. So, now I say, I want a very low-cost, low-power, but I still want low bit error rate. And then what people have done is let's slow it down. Let's use the 200G performance in the VCSEL, but then actually run it at 100 gigabit, right? So, now you have an excellent, excellent VCSEL that gives you a lot more performance. And if you use it at half the speed, you really get much better bit error rates or the signal-to-noise kind of ratio or relative intensity noise drops, as well. So, that's 100G. So, but you need more lanes, right? So, to get to 1600, you need 16 lanes of 100. And then recently, something came out called micro VCSELs, and that's going even slower, down to 50G. And then they call it NRZ, so instead of PAM4, which has four levels, 0, 1, 2, 3, now we go back to the original NRZ, which is 0 and 1. So, now you have the use of the entire 0 to 1 signal-to-noise, which again reduces your bit error rate. But you need more channels. So, you need 32 channels at 50G to get to 1.6T. But all three we are shipping, and all three are in demand by customers, depending on whether they want what they call fast and narrow, like 8 by 200, or they want an LPO, linear drive, no DSP, low power, and that would be a 16 by 100G. And then finally, if they want really, really low bit error rate, very, very low power, then they go to the 32 by 50G NRZ, which they call slow and wide. We tend to call it fast and wide, and then faster and narrower. But you know, when you're in the high-speed interconnection, we try not to use slow in any of our marketing.
