Silicon Photonics And The Future Of Ai Scaling John Bowers

TITLE: fjQ3Yorw-Ps CHANNEL: Unknown DATE: ---TRANSCRIPT--- We all use AI so much more in our daily lives than before and I think they’ll just keep expanding. And so I think that’s the most important thing to work on for futonics is new architectures of data centers that really are optimized for AI. The problem data centers have is you have a million processors even at a low failure rate. Multiple CPUs andor multiple transceivers are dying every day. And so for things like these, developing some AI model that takes a lot of processors and they all have to be there the whole time, you need to be able to switch out failed units. And that’s what the optical switching can do. That’s the problem with doing data centers in space is that failure rates are not zero. If you have a couple processors, fine, no problem. Just have a backup and no big deal. But if you have a million processors, it’s a million times more likely to have a failure. This week, Misha and Shinguay spoke with John Bowers, professor at UC Santa Barbara and one of the leading figures in the development of silicon phetonics. Over a career spanning Bell Labs, UCSB, and numerous startups, Bowers has helped shape many of the optical technologies that underpin modern communications. Today, we discuss the evolution of optical communications, the path from laboratory research to billion-dollar products, and what the future of AI infrastructure might look like when data is carried by light. Please welcome John Bowowers.

Well, for me, I mean, data centers and things have progressed very rapidly and there’s very complex CPUs and GPUs, but they require tremendous amounts of data, right? Tens of terabits uh for a modern chip and you just can’t deliver that with with copper and the losses are too high, the power is too high. So, fetonics really solves that problem. So I think data centers can be much more efficient because if you provide enough connectivity between these very expensive elaborate processing chips then they can make them much more effective. So I think we’ll see great advances with picks integrated in the package with electronics in the future. Maybe if you could tell us what is the uh current state-of-the-art or what is primarily used in data centers for communicating between the nodes. Well, today, you know, adjacent CPUs or GPUs in a like on a shelf or in a rack are all connected with copper. And the problem is that as clock speeds have gone up because, you know, a modern chip now is a 100 terab chip, a switching chip. And to get 100 terabytes on and off that chip, it’s a very dense array of of copper lines. And the speed is getting increasingly higher. So now 100 gigahertz or or beyond. and copper losses at those speeds. They’re typically not really waveguides, right? They’re just copper traces uh is is very severe. So you might literally have 20 dB of loss over just even a meter of of copper. So phetonics as you know the loss is the same whether you go a millimeter or 10 kilometers after that there’s some increase but uh it allows us to now connect not just adjacent uh CPUs together but ones across a data center and literally today you know there are million of them in a data center and they’ll be spaced at kilometers apart and so you don’t want to drive these capacitive copper lines because you know the power required is proportional to the capacitance uh that’s just very inefficient. Whereas optically the you takes a certain amount you know one EV to generate a photon and whether you go a millimeter or kilometer it’s it’s the same. So and then where place the the fiber optics in this as interconnects. Well, I mean today uh racks are all connected with with fibers and so you know the top the rack has a top of rack switch and again there may be say a million processors and say you know 100,000 top rack switches well less than that say 10,000 there you know there’s a bunch of CPUs in a given rack but then all those racks need to be interconnected and those are done with optical fiber today with lots of different switches aggregation switches and then spline switches above that and then there’s an interconnect out to the to the outside world. So when you make an AI request that comes in over the you know external internet down through that uh first switch and then down and the task gets assigned to some number of processors but increasingly you know data centers you you talk about warehouse scale computing really I mean the the whole warehouse is a computer right they’re all interconnected and working together it’s not just uh that you have a warehouse full of computers in fact they’re all working together so there’s a tremendous amount of data interconnecting them. Now it’s not so much just you know you send a request in and it comes back from one processor. No, it’s rather a whole lot of processors are are working together and that’s where photonix really has value. So the the other big revolution that’s happening now and Google is leading this portion of it is that top level of spine switches may be all optical and that allows you to not have to you know it’s really inefficient now to have OEO switches optic electronic optical so you have to detect the light you do your switching or do whatever your memory whatever you’re doing and then regenerate it on another laser beam if you can switch it optically it’s you know with only a couple dB of loss that’s much more efficient, much higher capacity and much cheaper uh to do. What originally drew you into physics which you studied as an undergraduate and then uh during your masters and PhD? Well, I had great physics and chemistry teachers in high school and so I started as as a joint major of physics and chemistry. Uh but I quickly realized I hated organic chemistry and so I just focused on physics. And so I originally was in high energy physics at the University of Minnesota, worked at Argon National Labs and Fermy Lab as a theorist or Oh no, no experimentalist. Um uh and I was I was the dumb portion of the you know I worked the night shift and did a lot of soldering no not much intellectual involvement. So what was the project you were working on? Well, the first project was at Argon National Labs where they have polarized uh proton scattering. And the nice thing about about Argon is they had these these polarized sources which typically does doesn’t exist. And and uh so that was interesting. In fact, the first my first paper was all alphabetical. I should have been first author as it turns out, but I was an undergraduate. They couldn’t have that. So, uh there’s one other author who really deserved it. I mean and then after that it started alphabetically with my name. So uh it was kind of funny but yeah it was it was a great opportunity for me. I I learned a lot and uh yeah and um later on why did you go into photonics specifically especially coming from high energy physics? Well, first I worked in solid state physics at Stanford and uh Gordon Ko was my adviser and and it was actually primarily more ultrasonics and and uh uh signal processing but there was a great laser group of physicists there you know Bob Byer and uh Steve Harris and others and uh uh when I went to Bell Labs then well first as a postoc then I worked with John Shaw doing fiber optics which you know 1980 was was brand new and yes and Stanford was a real leader in single mode fiber optics most of the world was still doing multi mode and uh so that got me hired at the labs who started that initiative from Stanford to switch to single mode fibers I think that was John Shaw’s vision from the beginning he was making sensors gyroscopes back then and you know the kind of fiber gyroscopes that eventually were deployed in in airplanes right and to make a low-noise sensor you had to have just a single mode multiple modes modes would would give you multi mode noise and limit your ability to sense rotation. So, you know, they were sensing, you know, 0.1 degree per hour sorts of angles and so it worked very well. Um, I see. So, they’ll try to reduce the noise, right? Multi mode noise would limit it. Yes, that’s right. And you just touched upon Bell Labs. So, Bell Labs produced an extraordinary concentration of people working on 35 semiconductors and opto electronics during that era. uh you were part of the broader ecosystem around people like Herb Krummer and Larry Cauldron who later helped establish UCSB as one of the major centers for compound semiconductor research. Uh what was special about that community intellectually and how did UCSB evolve into such a strong hub for opto electronics and semiconductor lasers? Well, that was Herb Cromer’s vision, right? I mean Herb was a tremendous physicist and you know he got the Nobel Prize for heterosructures probably more specifically inventing the double heterosructure laser which is what allowed fiber optics to happen right it wasn’t CW until that occurred but he also worked on the heterogunction bipolar transistor and was you know just a leader in just heterosstructures in general so his vision was to compete with Stanford and Berkeley and other universities in in California we had to do something So we weren’t doing silicon at all but rather focused on 35s. So the ironic thing about that is now I worked very heavily on silicon but actually I the initial part of it was with Herb and uh he had a lot of ideas and and in the end you know we’re getting gain from 35 materials but all the advantages of silicon so it was okay from his point of view but his focus was do something different than the rest of the world is doing. And you know back in when he arrived in the 70s you know everyone was focused on silicon no one was doing gallium marcenide or or anything else for transistors or photonics to be so was there like a competition with einhovven which is also known as a 35 place it is Idahoven is a very good photonic integration I would say they’ve remained much more focused on Indian phosphide integration whereas we’ve really and most of the world is I think really now pushing towards silicon photonics. So Bells gave you a front row seat to early fiber optic systems. Um what problems from that period most directly shaped your later conviction that every serious pick ultimately needs on chip gain? You know when I arrived in 1982, transmission systems were just at 50 megabit per second and uh you know the world had just gone through the transition from microwave transmission to optics and so they had to solve you know good fibers and then good lasers and detectors but the systems were very 50 megabit was a lot for a microwave system but the need for higher speed higher capacity systems was very evident and so I was lucky to get there at the beginning and we worked on higher speed laser and modulators and detectors and and during that time the first two gigabit systems were demonstrated then four eight 16 gigabit systems and that’s where it kind of plateaued because then uh everyone started doing wlink division multiplexing so by having multiple colors you don’t need and to go any faster and at that point about 1985 I think we did the 16 gigabit system that was really pushing electronics at that point in terms of amplifiers and all everything else you needed. So it’s easier just to then go to to multiple colors of light. Mhm. So those when you say like this 2 GBs per second or 16 Gbit per second, is it per channel or how is it defined? Well, at that point there was just one channel although actually they did have originally they had 8 micron transmission and then they added 1.3. So it was sort of WDM multiplexing but um not really that really didn’t happen till later uh when urban amplifiers were introduced and commercialized then you could now amplify all of the optical signals at once and so the cost of the system you know just if you had eight wavelengths it was 1/8 as much basically so you could really increase the capacity very quickly and what was the adoption of the technology is like to connect the continents or somewhere else certainly a lot of it was domestic but a lot of the focus was also intercontinental. So the first TAD 8 was, you know, 1988 and and just recently decommissioned actually. Um and so all the focus then was you know reliable lasers, reliable uh interconnects and uh so ironically the way they solved that because lasers were not considered reliable enough, they had a switch in. So they had four sources and a switch and it turns out that I don’t think they ever really I think the switches were less reliable than the lasers were. So uh and when you say lasers, this are like semiconductor lasers or and then mechanical MEMS switches or not even they’re mechanical switch like a relay kind of switch but moving fiber. Yeah. Yes indeed. They’re not they’re not good. So and you started in highspeed semiconductor lasers and mode locking devices before silicon photonics became the central story. So what did that earlier laser work teach you about what communication systems actually need or reward? Well, they always have to be cost-effective and so uh you know it’s important to to work on something that if your if you your vision is successful can be successful and and adopted and uh so we worked you know in the ‘9s at at UCSB on on Vixels in particular and Larry Cauldron and myself were both working in that area and those became widely deployed. Uh but at some point then speeds get too fast and there’s more need for an integrated circuit and so you typically need a modulator and detectors and other amplifier elements and that has really driven the transition towards inplane photonics rather than vixels which operate out of plane and uh and in the end um you know silicutonic transceivers are now literally a thousand times more reliable than pixel transceivers and so that really matters tremendously. ly in a data center because if you have a million processors and a million interconnects the just go through the rates it’s the failure rates of vixel transceivers really meant they had to replace multiple ones every day. Yeah. So if you can make that a thousand times better then there’s that’s the big advantage. So again, that happens because of integration. When you put a lot of things together, lasers, modulators, detectors, and maybe only one fiber optic interface, that’s more reliable than having hundreds of fibers and because it’s again that often limits the yield and cost. Yes. And reliability. And during that time at Bell Labs, did you already commercialize some of your work? Was that already happening um back then? Well, a little bit. Uh so right before I left I worked with a lot of other people in research Rod Alfres and Joe Campbell and others on different aspects of the first at that point was going to be a 6.8 gigabit system. The previous one was a fourth of that. And uh so we’re all working on Rod was working on high-speed modulators and Joe on high-speed receivers and I worked on uh lasers and and and the and the receiver. And uh so I was very proud of that work. And when I soon after I got to to UCSB, they killed the whole project. And uh why the reason was the the earlier system, the 1.8 8 gigabit system, a fourth of the other one was the most successful product in Bell Lab’s history. Why would they undercut it? And so that enabled Nortell to be successful because then they came out with the first the data rates changed a little bit. they went to sonnet and so it became 10 gigabit but basically was four times the original one and uh the labs was then behind and and that’s what allowed Nortell to get very big because what is fascinating um when looking at your career is that you’ve consistently worked in this pers quadrant where you can invent new devices that can then also be commercialized which I think as a scientist and inventor is the perfect place to be Do you think it’s still possible to do your research there or is it Oh, absolutely. I mean, you know, take any field, you know, like just new materials, right? So, you know, the with modern transceivers and and data centers that, you know, the you know, kilowatts of power coming out of a single rack, you know, thermal connectivity matters and there’s recently been some breakthroughs in high thermal connectivity materials. So, yeah, I think everything always starts with materials, right? And then but I think it helps to be focused on if this works if we can find this higher thermal conductivity material it will solve real problems and I think that’s happening now. You’ve worked with so many different materials. Do you have a favorite one or is that like asking having a favorite child? Well silicon is my favorite. Um clear answer well simply the reason is that you know CMOS is so highly evolved right the the you know as they gone through all these generations down to you know roughly one nanometer gate size the process equipment is so advanced and uh that it it really makes sense to take advantage of that and then all the packaging for you know 300 millimeter wafers and testing is so advanced that like India phosphide picks just can’t compete because there are too many gener generations back. So I really saw this when we worked with Intel because they were very careful about every making every process step very solid and so it worked every time and the process control was just unbelievable to to what I would see at UCSB for example and TSMC’s gone you know a whole you know the same direction. So if you make arrays of rings for some high speeded uh transceiver the this the the the resonance frequency of each of those rings is so tightly controlled in CMOS literally 100 times better than you could get at a say UCSB cleaner which has deep UV lithography so it’s pretty good but it’s many generations back so you know if we can find ways to get whatever interesting materials or whether it’s plasmonics or something else on silicon then you can really commercialize I think everyone would uh scream if gold came into maybe maybe other metals. Titanium nitrate. Yes. And you also started teaching entrepreneurship at UCSB in 1992. Why did you decide so early on that students should learn patents, products and company building alongside papers and research? Well, you know, I’ I’d been at at Stanford and that was uh there was a lot of startups coming out of Stanford even at that period in the early 80s. And when I got to UCSB, it was very academic and people were very focused on publishing papers and and and all the professors including myself were sort of focused on making clones of oursel and in 1991 there was a recession and uh people were some people were losing their jobs. And it was interesting to me because one student was brilliant uh was fired and or laid off and the reason you know is he didn’t really contribute. He wasn’t essential to them the company making money whereas another student who tried three times to pass the screening exam and didn’t. He failed each time and left. He became CEO of a company in 1991 and the difference was that company was under stress. So they they fired the CEO and a couple a bunch of other management people and they promoted this promising young engineer and I went okay this is interesting to me the the contrast. So yes I thought important that students learn management skills which you know it’s a lot easier than quantum mechanics but probably more important to the bottom line and also just think about what makes the company successful. What does the company really need rather than how do I get my ne next nature paper or something like that. I’m really curious how do you run your lab comparing to like maybe standard academic approach or what you’ve seen typically? Well, I try to have a strong group where everybody helps each other out because you know when you’re making photonic integrated circuits, there’s a lot of different tools that are needed and and it’s hard for one student to just be up on everything and everything. So, if you work together, things are better. Um, so, uh, do you put like two people on the project? I almost always do. Uh because too often probably in graduate school, you you may have experienced this as well. I certainly did as well. If you work by yourself, you can keep doing the same thing over and over again. Yes. And and you’re not all a sudden you lost six months and you didn’t really make any progress. Whereas sometimes just the act of explaining to someone else what you’re trying to do and what the problems are, even if they don’t know the answer, just the fact that you articulated it makes you go faster. And so I often well certainly have a senior person, a junior stu student and the senior student will be explaining to the junior student actually figure out what the problem is or maybe the other one will challenge them and say would you know did you really think about this or whatever. So yeah I think pairs or groups are always good. So nowadays for the um research groups in integrated photonics is this the time where you can delegate the fabrication to like foundaries for example like emphatonics get this multiple MPW um and then just focus more like on design and characterization or you still do fabrication uh on site as well? A little of both. So I mean if the founders can do everything you need then fine then that’s great and and you just have to do the design but too often to do innovation you have to introduce something else into the process and it will take them longer to do it and they may not be willing to do it until they see the value of it and so in the case of AIM we’ve had a good collaboration with them and what we our aspect was was developing an epitaxial quantum dot process to to grow on their wafers and then integrate that with the rest of their process. And so you know we do that in RMBBE machines and and uh we can cycle very fast right from growth till through processing laser maybe a week whereas for them a typical foundry to run a full process is 3 four months and and uh so you get a lot more cycle times you can learn a lot faster and same thing with heterogeneous when you bond 35s you know we did that initially at UCSB and again we could turn new devices literally every week and we transferred it to Intel and they do a much better job of it than we ever did, but their cycle times were a lot longer and so all the early stuff we did at ECSB with their involvement and then we transfer the processes over. So is it already released as a product by Intel? Yeah, they’ve sold uh more than a billion dollars worth of of uh transceivers, heterogeneous transceivers. Initially the focus has been 100 gigabit uh CWDM and as well as four fiber ones but now 200 400 800 gigbit. So um so yeah it’s been it’s been very successful. I think maybe like for broad audience if you could explain this uh heterogeneous transceivers how do they operate and why why do we need heterogeneous integration? Okay, so we like silicon because silicon is a very low loss waveguide and you can process it on you know obviously in in CMOS foundry so very high volume and that’s important in the case of Intel’s product they went from their their first 100 gigabit transceiver to a million transceivers in in a year and you know indie phosphide foundaries couldn’t scale that quickly right and that’s the advantage of that whole CMOS process so silicon technology is great and silicon waveguides are very good but silicon has an indirect band gap and it’s a very inefficient emitter one photon for every million electrons you inject. So whereas a 35 material would be 90 some percent efficient and uh so you need to get 35 you need to get some direct band gap semiconductor on the silicon and uh 35s tend to be very efficient but they don’t work well and they’re not reliable if there’s defects in them and so typically you need to grow them on a native substrate like indium marshide phosphide lattice match to indium phosphide and then we bond that to the silicon and remove the substrate. And at that point, it’s just a one micron thick layer, but it’s a highly cryst layer that’s defect free. And then we can do all the other processing in the CMOS foundry and and make lasers or modulators or picks or whatever. Um, how do you do bonding? There’s a few different ways to do it. Um, so many people do BCB bonding where you spin this polymer on the on the surface and then this kind of sticks it to to the other material. We haven’t done that and Intel doesn’t do that because the thermal resistance of that layer ends up limiting how much power you can get out and and the performance of it. So we do direct bond. So we we put an oxygen plasma and activate the surface and then we bond the activated 35 ox oxygenated surface with the silicon oxygenated surface. Do you need to like press it hard or you do and you know in the early days that was a big problem for us. would take pressure and a kneeling for eight hours at a time. Y and Intel said, “Okay, no way. We need we can’t have this machine sit here for that long. It needs to be minutes.” And they we D Leang and my group really focused on how do you make faster, better bonds. And uh u Intel then took it even further. So they they do it very quickly. So when they bond material, they bond lots of different uh chiplets basically because you don’t a lot of the pick doesn’t need 35 gain. It’s it’s resonators and modulators and a lot of multiplexers are very big and and and fiber coupling is very big. So they only put the 35 where you need it because it’s a relatively expensive material. And so they’ll they’ll take a handle wafer and put literally 5,000 chiplets on it and flip it over and bond it all at once. Now, the alignment doesn’t really matter because all the lithography steps occur registered back to that 300 millimeter wafer. You just have to make sure there’s 35 material everywhere you’re going to need it. And so, if you line it within tens of microns, that’s fine. And what are the requirements for uh for the material? because I understand there’s always defects like dislocations and how do you guarantee that um like the surface is good enough to for that bonding process. So when you grow on a you know lattice match substrate like on indie phospide or gallium marsenide then you don’t get dislocations or at least the dislocation density is very low 10 to the 4th per square centimeter. If you try and grow it directly on silicon, the defect density is typically 10 to the eth per square centimeter. And in general, lasers don’t survive long. They’re not very efficient at that defect density. It does turn out though that a couple of things are are reliable in the presence of defects. One are quantum dots, indium arsonsite quantum dots, because the high strain of the indium uh freezes the dislocation movement and so it doesn’t get any worse over time. Normally lasers are subject to things called dark line defect propagation that if there’s a defect they just keep growing and and and and quantum dot devices tend to freeze those dislocations in place. So that works pretty well. Gallium nitride for reasons that aren’t entirely clear seems to be efficient. It’s very hard and even if if there are a fairly high defect density you still get pretty good devices. Um but in general when we grow it on an uptax substrate and then bond it then we get this perfect crystallin material without any threading dislocations uh in it. One other advantage of heterogeneous integration is you can grow a wafer and now they’re growing on say 6 in Indian phosphogium arsonide then part of the wafer may not be very good. So when you dice it up, you do an inspection or characterization ahead of time and you know which die are good die and you only b m mount those onto the wafer and the bad portion of the wafer the part that might be lattice mismatched you don’t bond that you don’t use it. Is it just by mrology or you do like some optical tests? You do certainly optical scanning looking for defects in the in the taxi but also you do strain measurements and you can map that out. You do photo medicine measurements, you might map the whole wafer. And part of the wafer is dead on the wavelength you want, but part of it may be, you know, mismatch. And so again, once you map it out, you try and and work on the grower to to make everything perfect. But when it isn’t perfect, you only take the good portions. Does the epitextual process often times give you a gradient in thickness of the growing material? Is that an issue? That’s often an issue. You know, usually the wafers are spun when they’re grown. So the difference would be say center to to the edge there’ll be a difference not too much, you know, within a given diameter. And so again, the feedback is always to the grower to adjust the flows so as to make it more uniform. But a lot of times there are there are problems particularly on the very edge of the wafer and you don’t want to use that material. Then coming back to silicon, if silicon was such a poor emitter, why did the field spend so much time trying to build all silicon light sources? Well, I I think it was natural to try. Uh it’s not something I ever worked on because to me it was clear from the beginning that you had to have a direct band gap material. If it wasn’t direct band gap, it would never be efficient enough to be useful. And we’re seeing that in spades now with data centers, you know, they really require sort of just a few peakle per bit, you know, because when you have terabit transceivers, if they aren’t really efficient, they just take up too much power. You can’t get all that power off the chip. And uh so it has to be a direct band gap semiconductor. So if you could grow some combination of silicon or geranium and tin with strain and make a direct band gap, that would work. And so that’s what a lot of people did like professor Kimberling here worked very much on that but it hasn’t turned out to be possible or at least not at wavelengths that are relevant for data centers. So data centers have focused on 1.3 micron uh and telecommunications is based on 1.5 micron because that’s where the herbium amplifiers work the best. And so you know there’s been a lot of progress on germananium tin silicon strain layers uh at longer wavelengths say two and a half microns. Um but but that’s that’s an issue. And also just in general the the efficiencies are not nearly as high as you get with a 35 material. So you know if you need blue gallium nitrite is ideal solution. If you need near infrared gallium marshide is perfect and you longer wavelengths india phosphide is. So and in your group what’s the focus uh wavelength? Well we’ve evolved a bit. So we used to work exclusively at say 1.5 microns because telecommunications is all at 1.5 microns and then we’ve been working very much at 1.3 microns because data centers are all based there and then more recently we’ve been working at shorter wavelengths because then you get you know quantum applications and optical clocks and and things like that and so uh you can make very nice you know 850 or 780 980 nanometer lasers you know run 185 degrees C and very narrow line width and so that’s I think important for quantum and optical clock applications. How would you describe the main trade-off between heterogeneous integration and direct architectural growth on silicon? Heterogeneous integration works very well today. It’s commercialized. Uh it has very good performance. Um, in fact, I mean, in the early days when we first started doing this, we were just trying to make lasers on silicon as good as a native substrate laser and so make a DFB laser that would work as well. And we struggled for that, but that was always the push. But now we can actually make better lasers on silicon than you can on india phosphide or gallium arsenide or even gallium nitride. And that’s because the loss of waveguides on silicon is so much lower. you know it’s whether it’s a silicon wave guide or silicon nitride wave guide the losses are you know 100 a thousand times lower and that means we get much better line widths and silicon itself actually you know it’s very high thermal conductivity so if you say put a thermal shunt through the oxide layer you can now get higher temperature operation and so you look at Intel’s lasers you know they run at 150° C they put out you know 100 mills of power uh they work very very well so actually in many cases they work better than any native substrate device does. And what’s the line width of this Intel devices? Well, typically it’s 100 kHz. So, you can make something much smaller. And you know, by by integrating with silicon nitride waveguides, we’ve gotten down to one hertz line width. Oh, you’ve gotten down to one hertz. Yeah. Where’s the That is impressive. I I have heard from Jun Jun. He was for atomic clocks something like mini cards. Yes. Yes, with the silicon cavity. That’s possible. So narrow line with lasers are limited by thermal refractive noise. And so the longer you make the cavity, the lower the noise in that. And so right now the sweet spot is around 4 meters. So we can spiral it up something fairly small into a centimeter um and couple that to the laser and get down to sort of the you know well we can get we can get to 40 MHz Lencian line with high speed line high frequency line width for uh integrated line width we can get down to one hertz by locking to an optical cavity much like June does and that’s work in collaboration with NIST and and Carrie Vajala at Caltech and uh to get lower we need lower loss in the nitride the we’re we’re about you know a couple ten of a dB per meter right now and so if you look at multiply that by 40 meters then you know you’re sort of at you know a few dB of loss if you can get the loss lower yet then you can go longer distances and uh u and then you get even lower so we should be able the numbers say we should be able to get photorefractive limited about 1 millahz. So I think that’s where we’ll end up. And what’s the the the main driving force behind reducing the line width, reducing like the noise of the uh those lasers? Well, you know, for years people have always thought cementers are very noisy and you know, typical line width is a megahertz or a few megahertz whereas you know like a gas laser or solid state laser is you know kilohz line width, you know, orders of magnitude better. Um so what’s been exciting is by self- injection locking the laser to a resonator you can now get line whis that are you know just you know couple hundred hertz sorts of lency and line whis or or or lower one one one hertz in the best case and so it’s really uh changed people’s viewpoint and so whereas they would never use a similar laser because it was too noisy before now we could use it for a variety of sensors and or even for like say microwave photonics because the noise floor ends up, you know, you got a sit on a noise floor, you want to get a very large difference there. So, yeah. So, it’s it’s it’s in the end all a matter of reducing noise and and by having higher Q cavities, you can have lower noise lasers because you don’t have all the spontaneous emission from the additional gain you need. So, it’s all about making cavities that have a low as low of loss as possible. And could you give us a physical intuition about the self- injection locking mechanism because it’s surprising that this passive material in the end the silicon nitrite plays such a big role. Yeah, it is surprising and we didn’t really anticipate it. So that was fertuitous. Um so our our focus at the time the collaboration with NIST was to make very quiet microwave oscillators. So we had to make really narrow stable lasers and uh so uh we had gain regions and and then these highQ resonators coupled to them. And the idea normally when people do that they have an isolator to isolate the the gain region from the from the resonator. But the DARPA requirement was to very compact package. There wasn’t room for isolators. So we started just having soldering the two right next to each other basically. And so again, it was not it it doesn’t work for conventional locking like PDH locking. Um but but it works really well with self- injection locking. So the resonator has some back scatter to it. A very weak back scatter, but back scatter nonetheless. And that back scatter couples back into the laser and stabilizes it. So the laser line with literally over and over again we’ve done this starts at say a megahertz and then as you couple to a cavity it may reduce say 100 kilhertz. to make the cavity longer. It it goes down proportional to the length of the cavity and eventually we got down to 40 millhertz is the lowest laurencian line with that we’ve seen and so um it works really well. So DARPA really gave you a gift with that. Yeah, that was lucky the fact that they forced us their DARPA hard metric was very small package to have the whole uh system in. Now in that case they were looking for we had to do frequency doubled and you know octave and then frequency doubling to make an optical clock take a whole optical clock and put it on a chip and so there was a whole lot of technologies that had to get in that package it was it was very difficult and then for the back scattering mechanism so does it happen just because of like fabrication defects or you introduce some corrugations to control the amount of back scattering back into the laser cavity. So you can introduce corrugations and some people like Scott Pap have been doing that and and that works well. But with these really high Q resonators, I mean a resonators have cues of literally 400 million or several billion in some cases that you don’t need very much back scatter to stabilize the gain source. And so we don’t need to put corrugations in there. And the problem with the corrugations is it lowers the queue of the cavity. So we don’t do that. Um, but we try to make the very the most perfect cavity we can and get the queue as high as we can and there’s enough, you know, there’s rarely back scattering. There’s back back scatter from the sidewalls and and but the main point is, you know, if you have a queue of of a billion, then the energy in that cavity is so huge compared to the energy in the gain region, the the semiconductor portion that it just dominates everything. So what are the ingredients of maximizing uh the the cavity performance or like the quality factor? So they’re going to billions or how high can you go? So that that’s a case where silicon really works well, right? I mean if you look at waveguide loss on 35 materials losses are you know a db per centimeter is a very typical number. Um but silicon dioxide you can grow at very high temperatures very pure. silicon nitride very pure very high temperatures and and then if you do it in a camos foundry the lithography is so good that the roughness is very small and you don’t get scattering off the sidewalls. So then the main loss mechanism is hydrogen because often you’re using methane or uh you know to to and other hydrogen containing compounds to to to synthesize it. And so if you either use chlorides um like silicon tetrachloride things like that or just anneal it to drive the hydrogen out uh both of those work. It’s okay. Nitty-gritty details of the fabrication. And I also assume it’s uh the amorphous nature of the silicon oxide helps in this way or probably yeah but mainly just that you can make it incredibly pure, right? And you know that’s what the whole gate work for for CMOS has been is is to get very pure silicon dioxide. I see. Yeah, that’s incredible. Yeah, it is. But in retrospect, it seems like that for the 2006 um laser paper, it wasn’t even so much that you demonstrated the device, but there was this shift in thinking about the architecture of how these lasers should be assembled with the micro ring um cavity. Yeah, I think you know it’s like any field you know once there’s a demonstration and people realize it can be done they may not know how to do it but now they say okay you know we can try and replicate that and so I think it did change opinion of a lot of people and uh because it you know in in in that period around 2005 there was sort of fierce battles of all the people doing indie phosphater picks or gi marsenide picks questioning whether you’d ever use why would you ever use silicon? You know, it’s a it’s a lousy modulator because it’s not electrooptic. It’s a lousy laser because it’s indirect band cap and uh it’s not c it’s central symmetric. It’s not, you know, there’s, you know, there’s just a lot of problems from a photonic point of view and and so we made very good lasers and we started working on modulators, but other people uh certainly Intel and and uh Graeme Reed’s group and and Mahal Lipson’s group at Colombia and others made very fast modulators and and in spite of the fact that there’s no electrooptic coefficient, you can inject carriers and make at that point, you know, one or 10 gigahertz modulators, which was sufficient. That’s where the field was at the time. Now people are making modulators in silicon that are at literally 400 gigabit per second and that’s just was inconceivable at the time. So there a lot of different vectors that had to happen to convince people to switch to silicon photonics. Also speaking of uh let’s say lasers and integrated photonics chips. So typically when they see the papers people let’s say inte integrate photonics they usually like show the chip but then no one really knows what the infrastructure you need to run it like they for example if it’s like a passive device they like oh we we somehow have a laser somewhere on the shelf but so what is needed in terms of infrastructure for your chips I mean I understand that laser is already integrated but what is needed to support it integration helps you out because you don’t need to be aligning chips or having fiber. If you have fiber in there, that has adds noise, right? Because there’s there’s acoustic noise, there’s there’s thermal noise with that vibrations. Um, and so integration makes the need to stabilize easier. You can you can integrate say the the cavity or the the the wavelength locker or whatever or the detector to to monitor what it’s doing. So, in the old days, you’d have indeed spectrometers and all these other power meters and things around it. Now you can put all those things on the chip and it makes it a lot easier to use, but you still need power supplies and things like that. And so again, modern modern devices, you bond the electronics to the pick and so you you can easily have a electronic IC that puts out say the 20 signals you need to make a very stable transceiver or whatever the device is. So you once described automated packaging and wafer scale testing as critical for the future of the industry. Has the field caught up to that view? Yeah, I mean that was my opinion. Um when we first did the first uh presentation of the Intel pick, uh Pat Gellzinger who was CTO at the time then was CEO of Intel, he really made that point that the photonics industry does does not understand what this means. And uh he said you know they’re used to it’ll take a day to to uh characterize a particular transceiver. He says you know we do this wafer scale we do the whole wafer in in in 8 hours and we fully understand everything on that chip and we know which ones to package and and it’s just it’s a revolution. So he was really kind of the main proponent of that concept and but I think it’s ab very very very true and uh so um and you know like companies like Nexus do that in spades they’ll a fab will finish during the day and the next morning the entire wafer is is characterized. Oh so you can do that if you have onchip lasers you know how they’re all the whole thing will work. So you can charact you know you can do loop back and yes it may be 100 gigabit transceiver but the laser the modulators detectors it’s all there you can test how well they work right there and whereas if the laser is not integrated then you can check say the pick by injecting light and and detecting the light externally but when you actually connect with what will end up being the final laser it may behave a little differently or there may be reflections we have experience that yes yes so that’s the advantage of laser integration. So what’s actually like the footprint of of the laser because I understand it you have internal cavity and then external one and then so what’s what’s the total space it takes on the chip? It depends on the application. So a typical laser with say a modulator and a detector is probably a millimeter long by 100 microns wide and uh maybe maybe could be longer than that but on that order. Um now if you integrate a very high Q cavity say a 4 meter cavity that’s now going to take a certain area probably like a centimeter by a centimeter on that order to do that. So it becomes the narrow line with lasers become a lot a lot bigger. It’s like the narrower the laser the bigger the space. That’s true. That’s true. I see. Do you think the primary bottleneck today is still device physics or has it really shifted toward packaging, testing and supply chain infrastructure or are these just two different ecosystems that mutually help each other? Because a lot of the photonix ecosystem still looks like, oh, a grad student went to clean room to make something small volume but maybe very um very novel but not necessarily super scalable. Well, certainly I mean packaging is incredibly important, right? It’s it usually dominates the cost and and the performance of the overall system. Uh but there’s still need for device innovation and you know again in the case of successive generations of transceivers. So now we’re sort of at 3.2 terab people are integrated onto a chip. You know to go to six or higher you typically need faster modulators and faster detectors. you need more linear devices to get say PAM 4, PAM 8, uh or you may need devices for coherent communication, you know, 256 QAM or things like that. So, we haven’t by any means reached the limit of device innovation. Both sides are very very much needed for for progress. So what’s interesting with the ultra narrow line with um laser is that previously we were thinking of um of semiconductor lasers as these very uh messy systems and you reach this regime where it’s essentially the same as precision metrology that you usually do on an optical table and all of a sudden you’re also sensitive to environmental influences and all sorts of noise. Was that was that a shift in thinking lab infrastructure to kind of bridge this and get much much cleaner with characterization and noise um characterization? Yes. Well, we learned a lot from Nest, our collaborations with, you know, Scott Diddhams and that that whole group. They’re the experts at at you know, phase noise measurement. And uh uh it really does help to get rid of optical fibers because they are vibration sensitive and and just with temperature changes day to night then you see yeah you know polarization changes they’re good microphones. Yes good sensors that way. Um so I I think integration really really helps there. Uh and then again keeping everything within the same thermal environment helps as well, right? It’s harder to stabilize when you have multiple lasers and things, temperatures changing and in the room or whatever. If it’s all in the same pick, then it’s a lot easier to do. So that already gives you some advantage in face noise. Yes. And then usually the lasers at what um bandwidth do they operate? You mean like how fast can you tune them? Yeah. Yeah. So it depends how you tune them. So if you change the semicond element typically it has bandwidth of you know tens of gigahertz and uh so you can make a you know you can integrate a high-speed phase modulator there and that’s important for applications like LAR right where you want to chirp the laser for an FMW system very rapidly and uh uh so that that that works well if you thermally tune the resonator that will tune the frequency as well and that obviously is a lot slower sort of you know kilohz sorts of speeds Yeah, but that’s kind of nice then in terms of um noise uh performance because uh most of these applications like LA that are commercially relevant, they just bypass all of that messy low frequency noise. Yeah. Although again you know I think pics will have a big impact on lighter because you know what what today is sort of a very not integrated in many systems like an FMCW system it’s not very integrated and uh you got fiber coupling and things like this you can put almost all the elements if if not the entire element on one chip and so for instance and has been is a company that that is working on that and uh so is typically heat management an issue for this kind of lasers and how do you fight with it or solve the problem? Yeah, it is um you know particularly as you get more and more devices integrated together, right? So um you know like the 3.2 terabit chips there’s a yeah there’s a lot of laser power and a lot of amplifier power, modulator power that you have to dissipate. And the problem with silicutonics typically is you got a silicon layer but then you’ve got this box layer, the silicon oxide layer and then the substrate. And so that unless you put a heat sink on the top that could that can limit your thermal performance. Silicon substrate is ideal and even the silicon waveguide layer because it’s very high thermal conductivity. But there are ways around it. You know, you can put holes in the oxide and fill it with polysilicon or fill it with metal or whatever. And so you can actually get a fairly good heat sink down to the substrate as well as putting a heat sink on the top of it. So we touched upon this topic a little bit in the very beginning. Um, but coming back to AI and data centers, we recently interviewed the CEO of a company that wants to build data centers in space partly to get effectively free cooling, which um felt like a pretty good illustration of how extreme the thermal and power density problem is becoming. At what point did it become clear to you that electrical interconnects were turning into a fundamental bottleneck for data center scaling and that silicon photonix was becoming less a niche technology and more an architectural necessity? Well, I mean, it’s been true for a long time that that copper has been a limit in the losses of copper and the power to drive it because again, you know, if you look at most existing or certainly older Cisco routers or anyone else, you know, they have plugins at the end and they have that long copper line going back to the chip in the center of that printed circuit board and the laws are so high they have to regenerate the signal at the end, right? So you have to clock recovery and and regenerate it and that takes a watt per channel and that’s that’s a lot of power when you multiply by you know a thousand or 10,000 and uh so it’s been clear that you had to get photonics much closer to the chip and get rid of all these losses and so um whereas today indeed all the all mostly today all of the servers in in Iraq are connected with Ethernet cable copper Uh that’s very I think in any new build that isn’t going to be true anymore and it’ll be photonically integrated and it just saves you so much power not just the power of the actual signal but the fact you don’t have to regenerate it. So are we right now just at that junction where we are switching over from copper to um photonic interconnect within the rack. Yes. Yeah. Within the rack. Oh I I would have thought that that already happened. I mean I don’t know but it has happened to some extent. Certainly five or 10 years ago it didn’t happen. That’s really interesting. Yeah. But now it’s really getting very close. So like Broadcom and others have announced co-ackaged chips. So you know they’re they have a 50 terabyte chip where the interfaces are optical in that same package and that’s required a tremendous advance for silicon photonics because now you have to make very reliable. You have this okay expensive chip with all those photonics around it. It can’t fail and it has to yield. And so you can do onwafer testing before you dice up the silicutonics wafer but then you know you can’t have any loss of performance when you mount it in there and that’s required a lot of other innovations like detachable fibers right so if you look at how people assemble a chip high temperature processes maybe not very nice to optical fibers so you need to have then connectors that detachable fibers that you can connect onto that chip after it’s been gone through all the heat flows for solder and things like that. So the things like robustness and yield were those the main um bottlenecks to solve before it could make it into data centers. Yes. But but I mean I think they have been solved. I mean they are great. Broadcom is selling the OFC presentation this year showed pretty high volumes of co-ackaged optics. That’s really cool. But for some reason I thought that it already happened. You should go to data scient. So just maybe to give like a bigger picture about uh interconnects in data centers or like in hardware. So there is uh onchip interconnects then there is between the chips and then there’s between the racks or maybe if you could so comment on that. So, you know, initially it was pluggable optics on the side of the board and then we went to onboard optics so that you’d have these photonic transceivers around the switch chip and that’s certainly where we are today. Yeah. Um the next step which Broadcom and Intel have done is to have the phonics inside the chip. Um and then you might have multiple GPUs that are talking to each other through these fiber optic interconnects. And then um you know the next step is on the chip itself communicating different portions uh optically. So you know you’ll have a trans you know you you’ll have a a GPU on on that chip and there might be other GPUs integrated on that same chip and and may connect optically to them or it may go off chip. And so that’s the the goal we’re all trying to get to is even at 100 microns you’re better off sending it optically. you just have the same interface whether it’s 100 microns or 10 kilometers. So where is the biggest need now in the enabling like going to the next level? I think just higher levels of integration because you know the the field keeps advancing right. So I I mentioned that Broadcom had a 50 terab chip that that has optics in the package, but now you know they’re working on 100 terab they’re selling 100 terab chips and in two years it’s going to be 200 terabit chips. And so you know the it’s kind of amazing how fast the field continues to progress. So you know very few fields you’re you’re doubling either your speed. I mean how do you get the ever higher capacities? you either have more lanes which takes more volume more area or higher speed devices and so or more colors of light more WDM multiplexing so I mean there’s just a lot of pressure on the whole ecosystem to keep evolving to get higher and higher and then of course you know if even certainly 10 years ago it was inconceivable you could make a pick that would transmit three terabytes that would actually yield in any anything greater than like 1% right you had to have all the individual yields of components be that high that you can actually do that in in volume and at high yield. But that’s now where the field is and that 1% would be the nature paper. Yes, indeed. Indeed. An existence proof, not something that was at all practical. Yes. Is there anything in particular that is still holding back quantum dot lasers from becoming widely deployed at data centers? Well, quantum dot lasers have a lot of advantages like they’re, you know, insensitive to defects and and uh so actually they’re very good for things like put it in space where there’s a lot of exposure or um you know nuclear tests. So um Sandia did some tests recently and published a paper where it’s it’s 10 times more reliable in terms of neutron and and and other things than the best commercial india phosphide lasers they had. So that there are advantages. The other big advantage of quantum dots is the lower line with enhancement factor means that they’re reflection insensitive. So normally we think lasers need this, you know, 40dB isolation. Um the sensitivity of a laser to reflection goes like one over the line with enhancement factor to the fourth power. So with quantum dots that factors rather than four is 04 or even 0.04. So it’s either 40 or even up to 80 dB less sensitive. So you don’t need isolators. So that’s important because you know if you’re combining lasers and picks on on the same chip, there’s no room to put an isolator in there and you don’t want to have a magnet and things like that. So that’s the advantage. Um, but fortunately these highQ, we’ve been talking about these highQ resonator couple devices, they’re also very reflection insensitive partly because the reflection back through the resonator u is a very good filter, but also because it just stores so much energy that the laser is not disrupted. Yeah. Uh like a just a little 1 millm long semiconductor laser would be. It’s really a surprisingly neat design. Yeah. Yeah. During the telecom bubble, companies like Kali were built around the assumption that optical infrastructures demand would scale explosively. Um, the technical vision was not necessarily wrong, but I mean it wasn’t wrong, but the economics and timing failed. Are there any lessons from that period that are being ignored in the current AI buildout? Sure. Well, I mean, everybody’s worried that this AI buildout will be just like the internet one in 2000, like with all the fiber and no data to send through it. So, you know, Cali was founded with the idea that you could reduce power and cost by sending the signal optically through there and not having to regenerate it with this OEO in between. And I think we’re just ahead of our too far ahead of our time uh at that point. And uh it there was this whole collapse. So, at the time there were literally 60 companies doing optical switching. Oh, wow. Of which Caliant was just one. And by 2002 there were only three left and I think Kellant had the biggest switch and probably the highest volume of of production and uh but still you know it’s been relatively small. So Google has which was Caliant’s biggest customer has now really deployed it in data centers for all the reasons we’re talking about and they’ve taken out that whole spine level of of switching and it’s just optical and so um that’s uh the biggest application of optical switching. So we were literally 25 years too early. Um but all the arguments we made you’ll now hear in a Google talk and and uh and it worked. I mean it’s it’s there’s no need to electronically regenerate it necessarily. So you know the problem data centers have is you have a million processors even at a a failure low failure rate you’re still replacing multiple CPUs andor multiple transceivers are dying every day. And so for things like these developing some AI model that takes a lot of processors and they all have to be there the whole time, you need to be able to switch out failed units and that’s what the optical switching can do. And so and that again was our original premise as well that redundancy and things would would be needed and an optical switch allows you to to replace a CPU that failed without having to send someone in and and pull a card. Yes. Or someone in a satellite in particular. Yes. Yes. I mean, that’s the problem with with doing data centers in space is that you know the space. Yeah. It’s you can’t replace things and and failure rates are not zero. So I mean if you have a couple of processors fine, no problem. Just have a backup and no big deal. But if you have a million processors, it’s a million times more likely to have a failure and and I’ll sign up for that job. A space AI data center technician sign. Might get a little bored over time. Um so from your point of view what what is the uh the main break breakthrough in integrated photonics you’ve seen over maybe like past year or over the five years of late there’s been I think tremendous progress at going to shorter wavelengths. So, you know, most silicutonics still uses silicon as the wave guide. And that works great for dataccom and telecom wavelengths. But if you want to work in the visible or in the near infrared for clocks or or things like that, then you need to use nitride wave guides. And uh it turns out you can actually make very good lasers by bonding 35 material on nitride. And u I think that’ll have a big impact. So if you know you can buy a commercial laser say from Optica at you know 780 or whatever but you know really only runs at room temperature it’s big and expensive and things like that but by making these lasers on nitride now they run at 185 degrees C. So there really isn’t a temperature con constraint and you can now do more interesting things, right? So a typical, you know, clock experiment probably needs modulators on there and and needs detectors integrated and things like that. And so I think the shorter wavelengths are are are the newest and the most exciting thing I’ve seen of late. How much power can they provide those lasers? Yeah, 50 100 mills of power. Okay. Because typically also I feel like in uh atomic and molecular optics you need you’re really hungry for power right so how how do you see that um picss can address that or like I mean the scar of lasers well the other advantage of of integrating gain on the chip is not just having a laser but then having amplifiers right so sure as we make ever more complex chips with more modulators and and and elements in there then there’s there’s a need to make up for the losses of those elements. And so amplifiers allow you to make things that are much more complex than if you didn’t have it. Just like electronics, if you didn’t have an electronic amplifier, what you could do would be very very primitive. And uh so I I think that’s the other advantage of having gain integrated on chip that you can make up for those losses and uh make the performance better. For people outside of the field, what actually is a frequency comb physically? Why was it such a revolutionary idea uh once people realized you could phase lock all these optical frequencies together? Well, you know, we know from 4A transform theory that if you have a bunch of frequencies and they’re all locked with, you know, identical phase together, then the 4 transform that is a series of pulses and the more lines you lock together, the shorter the optimal pulse will be. And so, you know, that’s not new. That’s been known for a while. I think uh with combs, you know, like Kirkcombs as you mentioned, if you have two lines and there’s nonlinearity, those will generate some indifference frequencies. And so you’ll get um you know, four lines out of it. And and those four lines will generate eight and and you get to some, you know, large number of of lines uh that are all well, it may be a noisy comb. And so the big push has been to make quiet combs. So if you have a lot of solutons, you know, these these pulses in the cavity, then it’s noisy and and uh what you need is a single soliton. So the problem with a lot of Kurcom technology was that, you know, you had to adjust the power and the frequency just right to get it into that single soliton state. And that was difficult and it didn’t always happen. So you know, you kind of needed the PhD there to turn the knobs and and if it didn’t work, do it again. Um, but the advantage of the self- injection locking and driving it hard enough high enough intensities to get comb generation is that you get a turnkey situation where the you just turn the laser power on and it goes merely into the single soliton mode. You don’t need to adjust the power, the frequency and it always happens. And that was really in in our case really led by Carrie Bahal at Caltech. And uh it works surprisingly well and I think that’s what enables it to be then deployed commercially because now that you have this large number of lines all exactly equally spaced all low noise you can now put modulators in each of them and uh get a very high capacity. So you know again today data centers are typically say gone from one wavelength to four wavelengths but soon six and 32 and then 64 and having 64 lasers all temperature controlled and power you know the spacing just right is very complicated and a single comb source really solves that problem. What exactly is a solidon in this context? Because like the first occurrence of that is someone riding along a river and seeing that wave that doesn’t change shape and then again a lot of people are just thinking of it as an abstract mathematical object. How do you think about soliton or what is it in photonics? Sure. I mean a soliton you know is a balance of the dispersion that wants to cause the pulse to spread out and the nonlinearity that is is is pulling it back together again. So just the right power, the soliton power, those two balance and the pulse is stable and uh so indeed it happens with water waves and tsunamis. Um and it happens on optical fibers and I think Lynn Molenour is the first to to observe that and uh uh what we have here is the same thing. The pulses are solons but there’s a train of of pulses which is what you have when you have a whole chrome of frequencies. So the the modes are solid but it’s a whole train of solutons. The field initially obsessed over bright solaton states uh but then dark pulses started looking extremely attractive from an efficiency standpoint. Uh what did that teach the community about the difference between let’s say beautiful physics and useful engineering? Well, they’re both important. Um but indeed you’re right. As you get more and more comb lines, the optical pulses get shorter and shorter. And so the end result is almost all the time there’s no power being transmitted and the peak power is very high. Whereas if you have a dark solon it’s basically on and there there’s this little dip and the more lines you have the narrower that pulse is. So indeed you get away from some of these high nonlinear high pro high power problems and uh get more perhaps microwave transmission if it’s on half the time that gives you the most power into the fundamental frequency you’re trying to generate. So for a lot of microwave microphotonic applications that is a preferable way to go. Mhm. And um one theme throughout the review on microchromes is that integrated photonics are slowly rebuilding the architecture of a precision optical lab directly on chip um reference cavity synthesizers nonlinear crystals and so forth. Did you always think uh the field would invol evolve into that direction or does did that came as a surprise once you had the really good lasers for example? Actually I have thought that for 30 years. So I’ve taught a short course at OFC for 30 years now and I always had a slide on microwave applications of photonics and uh just the fact that you know microwave cables are are big the connectors are big they’re expensive um microwave losses are very high thousand times higher than optical fiber so I have always thought that photonics would really in influence microwaves and that has now happened right I mean mostly other people’s work that you know optical clocks are now better than the Fast clocks are optical not microwave and uh uh I I have looked forward to this vision where microwave devices say from Ksite and others would have backplanes that are optical rather than cables as they are today. And uh and indeed it circling back to silicon photonics you know it’s it’s the difficulty measuring these very high speed devices at 100 gigahertz 200 gigahertz is you’ve got you know a synthesizer source here and then a cable and a probe and and then the other end you take it off there’s a lot of loss in there and at at 100 gigahertz whereas you should be able to have just an optical sensor right at the tip of that probe and that’s measuring the signal everything else is optical and I think that will happen we’ll see instruments that are that way. So maybe if you could explain how do you merge this two different frequency bands like one is in hundreds of terraertz the other one is like gigahertz range. Well, you know, you need to if you make a state a comb and you generate an octave of frequencies, then you can double a line on the low end and and mix it with a line in the high end and and that’s what allows you to get these incredibly stable clocks. Uh because indeed, you know, individual optical lines are are moving and they’re not they’re they’re not stable. But when you double that and mix it, that offset frequency has to be low. something you can measure tens of gigahertz. Yep. And if it’s a terraertz comb as high as a half a terraertz and and that’s the problem. So you need enough control that you can make sure that’s in the measurement region, but that’s being done now. And and u so I think we will see we’re fairly big optical clocks really be all on a chip and that’s huge, right? So if we all had an inexpensive, you know, chip with incredible accuracy, then you can do precision navigation. And so at some point perhaps every vehicle has has one in it. And uh because indeed, you know, to measure where you are, you need to know time. If you know uh if you know velocity, you need to know time to know how far you’ve gone. And uh you can build optical gyroscopes and and so I think that really will become important for you know autonomous driving in particular um or any GPS denied environment which is pretty important in places like Ukraine. So have you done any uh gyroscope on chip type of work? We have about uh five years ago uh a student of mine integrated maybe more 10 um the entire pick on a chip. So the delay line as well as the front end. Um and that was part of our motivation originally for working on very low loss optical wave guides and uh a company on Nello is now working on on that. So um you know there’s a lot of um you know fiber optic gyroscopes work very well and and they’re using airplanes and things. We we’d like to get that and we can get that on on a chip. The problem is that the sensitivity of gyroscopes proportional to the area enclosed. So if you make the loop pretty large diameter for fiber, that works well, but that takes up a lot of area in the silicon chip. But you could do that today. I mean, you could just have instead of a um dice, you you have like a uh ring of like a silicon ring with your 300 mm wafer for just put it there. You cut out. I mean then you may as well just use some fiber. So I think what Anel was doing is more integrating the whole front end all of the the lasers and the splitters and the detector onto a chip and then have fiber which you know works well. Yeah. So so it seems like we’re moving away from purely just telecom and communications based applications to many things that are quantum adjacent. Do you think that that’s the more relevant uh market before quantum computers become mature? Is that going to be a substantial field of application for integrated photonics? It probably happens sooner. I mean, I think quantum communication is something we can do today and maybe not as well as we need to or as high a rate as we need to, but I I think given study, we can make very good quantum communication that’s very secure. So I think that probably happens very very quickly. Same thing with sensors. I think you know magnetometers and things quantum works extremely well and and uh uh so I think that’s a big application and increasingly you know in GPS GPS denied environments you’d like to navigate based on magnetic field and so these quantum magnetometers I think are very interesting and will find big application. What’s your opinion about uh uh optical computing? I think it will happen because again you know there are some things like vector matrix multiplication that work well optically right you can do very high capacity very fast calculations of very complex large tensors and things um how about neuromorphic computing to just have different architectures to start with instead of foyman uh computing architectures yeah maybe something where photonix could could be of help. Yeah. I mean, the the problem has been it’s been like optical switching, right? We we talked about the advantage of optical switching, but it took 25 years until well, 20 years at least until it really became widespread. So, all Google data centers have it now. And so, it it is now deployed, but it took 20 years. And part of the problem was electronics kept advancing, right? The electrical switch chips didn’t just sit there and stay limited at at 100 gigabit per second, which is what it was in 2000. they just kept getting advancing, getting better and better and having more a wider range of of functions they could provide. That’s the problem with optical computing is it can do something very specific but it it may not be a general purpose computer that that silicon will do and silicon is not conventional silicon camos is not staying put. So I think it will u as a special purpose processing unit happen. Um are we happy overall about u like the optical switches uh or there needs to be more done like what’s the problems? There needs to be more done. I think uh indeed what they would like the the data centers would like would be ever larger radics right so if you have a a smaller switch say a 20 by 20 switch then you have to have multiple stacks of them and lots of interconnects connecting all these smaller switches if you make it a bigger switch then it’s always performs much better and has much more value so today mems switches are about you know 300 by 300 and they would sort of like a thousand by thousand and uh and so that just that’s a lot of hard optics problems and and yield problems to solve but I think it’s very doable and and will happen. Um the other issue typically today is that loss matters a lot because the margin of these transceivers is not very large say 12dB or 10 dB and much of that loss is taken up in connectors and and perhaps splitters things like that. And so if the switch has very much loss then it it requires them to have more expensive transceivers that have a larger margin. And so the typical requirement today is 3dB loss and that’s hard. You can do it with me switches, but with planer switches, it’s hard to get, you know, coupling loss in and out plus the loss on chip to get below 3dB. So, um, I think there’s a lot, we’ll continue to see a lot of advantages. The other thing is that the like the Cali MEM switches I’ve been talking about, those are dumb switches that all wavelengths go through, which is the advantage. You can then put hundreds of wavelengths through. That’s the good part. But there’s a lot of value in having wavelength switches and being able to, you know, have a transceiver generate 30 wavelengths and switch which wavelengths go to which ports at the output. And so I think we’ll see a lot of evolution there, more complex, you know, functionality basically. So do we also see like a role of inverse design because it there have been a lot of studies happening in like university labs. How does it propagate towards maybe products or industrial level applications? Yeah, I mean inverse design works very well and and I’m always amazed when I see a very compact high performance chip and uh so yeah I think you know to a much larger extent it’ll be deployed by by every photonix designer. Um so yeah it it it you know size matters and performance matters and and in inverse designs come up with things that you and I would never imagine right so I guess it depends on how flexible are the foundaries to to fabricate those things but the lithography is so good they can make it right and they can make it reproducibly so um how do you usually go about starting a company at what point does the decision fall that you would like to spin something out from your research well in the early days, you know, I if there was something I was really excited about, Terabit Technology was the first one. I took leave of absence and and with some students, we started the company and and uh and that eventually sold to Sienna. And same thing with Caliant. I took two years leave of absence to start that. Um but more recently, I’m much more just a coach, right? So, if I can help someone else be successful, then, you know, I’m happy. And so, you know, Nexus and Quintesscent are are more that way. So, again, good ideas coming out of the lab. And, you know, if I can help someone avoid mistakes, then then that’s that’s worthwhile. What are like the top three mistakes one should avoid if Well, the first one is running out of money. So, you know, I always tell CEOs, okay, you know, your main job is is not technology, but just bringing the money in and uh and getting the product out. And so uh I think those are the two ends of it. One not running out of money and the second is not enough product focus. And so you know knowing early on really identifying what the product is, what the specifications are and then focusing everyone in the company on on getting that out and uh so the successful ones are that way. Um the best example is a professor. So I was working with William Wang who’s in AI and he’s using AI for chip layout and the company’s called chip agents and uh tremendously focused. He’s always figured out what the product is from day one. Everyone’s focused on it and I think 10 of the top 20 IC manufacturers use his layout process now and it makes everyone much much more efficient. And so it really made me a true believer in AI and the value it has because all these different agents are looking at the problem from different aspects. One’s optimizing power, one’s optimizing area, one’s, you know, checking out uh faults and you know just it’s uh I I’m excited for when AI really hits photonics in a major way for all the layout problems. I think we’ll see a big advance in performance then. Do you have anyone in your group working on that as well? Have you dabbled in that yourself? Dabbled, but no one I mean uh no one’s currently working on it now. So any like plans? How do you plan to use it in your lab? Well, you know, I’m retired so you know the group is getting smaller. So uh um but yeah, uh Darkco Zebar was doing AI a couple years ago in in the group and but that he was kind of the last one. So you’ve earlier mentioned that uh there was a time where you were too early with um with an invention or with a uh concept. Uh is there anything where you you’ve ever been too late where you felt like uh well probably I mean the data center space I think uh uh you’ve got to have a lot of conviction that that you’re right in the end and and so I mean most silicon photonics does not have integrated lasers. So indeed you know that that has been my focus and Intel’s been successful and quitesscent to Dexus and and Open Light. Um so I’m uh we may have been a little early on that but I think as people need more complex functionality that will be the only solution. So I think it will it will happen. I think everyone will be doing it at some point. So the big switch will be when TSMC does it. If you would imagine that you have unlimited resources, what project would you uh commence? I think the most exciting and important direction right now is in is in data centers and in AI and uh new architectures of data centers that really are optimized for AI rather than for you know conventional applications. And so again, you know, we all use AI so much more in our daily lives than than before. And I think that’ll just keep expanding. And so I think, you know, the the need is going to be much greater in 10 years than it is today. And so I think that’s the most important thing to work on for photonix is is to make ever higher speed transceivers and ever more integrated with the electronics in that CMOS line. Um I think sensing is hugely important. Uh again, we’ve all seen the advance of autonomous driving and and I love Whimo and but those are big expensive LARs and things and I think u we can build integrated LAR and and and they’ll be cheaper and and more effective and you know I think it’s a tremendous advance right if if we all had the ability to to measure velocity of things coming at us in the fog at 300 meters away and you know it’ll be much safer and So, uh, you know, I think every car is going to have multiple optical lightars on it and and so sensors in general, whether it’s that or glucose sensing for, you know, things like that. I mean, you know, our watches have a lot of optical sensors on them now, right? And I think that’s going to keep occurring and and we’ll all be healthier for for that. So, if Bell Labs were rebuilt for the AI era, what capabilities would need to exist under one roof? The value of Bell Labs was partially that you had world experts in everything you needed. So like when when the switch came in the in the mid70s from microwave transmission to everything that’s going to be optical transmission, you know, there was groups folks the best optical fibers and and the best lasers and modulators and detectors and electronics and and you could go down the hall and talk to someone and get get an answer and and you can move very quickly. So, you know, having people that are good at at growth because again, a lot of innovation always comes from new materials that solve some problem, people are good at devices, people are good at circuits and and certainly software, electronics. So, um just you know, that’s what I would I would put together. And in that vein, I saw the recent call from NSF, which um wants to establish NSF labs. Uh do you think the US research ecosystem still knows how to support long horizon technological infrastructure projects? Is that maybe one step towards it? Yeah, I think it is. I think it’s an excellent idea. Um the the problem is so much research is short-term focused and it’s sort of individual small groups. I think there’s real value in getting larger groups of people to invent an area or to dominate an area and make a lot of advances. And I think other countries have been more effective than we have whether it’s you know Taiwan with displays or yeah, you know, Korea or whatever that they they’ll focus on an important area and and dominate it and and we have done that and we can do that. We need but you need longer term vision and uh so I hope we do that in the quantum space. I think we can and same with the AI space as well. I think we have a lead and we can really expand it. So uh I think the XLABS is a great idea. Min warned us before the interview that you are apparently incapable of doing anything in a non-competitive way. Is that an accurate characterization? And what’s your favorite thing to compete in? Building better lasers. Well, I think there’s some truth to it. I I won’t deny it. But when students come to me with an idea, they want to work on something. One thing I’ll some if I don’t think it’s really pushing as far hard as they should, I’ll just say so suppose you’re completely successful at this will anyone care? And and oftentimes the answer is well not really because it’ll take two years and by then the world’s gone on to something else and and so it is important to look at impact and I’ve always wanted students to have impact and if their research is successful that someone will care about it. So um yes because you can build the most beautiful thing and no one really cares and then it’s art rather. Yes. I mean science is good but science good science is important but hopefully it has an there’s there’s a reason people will value it whether it’s detecting gravitational waves or or quantum computing right there’s there’s value in that. So, do you see any parallels in your love for extreme endurance sports and building photonics like integrated photonics? I guess well there are similarities there in terms of like getting a PhD, right? I always talk to students there’s going to be a lot of nights and a lot of despair and and uh experiments going bad and and things like this equipment not working. So, you have to be really determined that this is important to you and has to come from within, right? And and you know, I’ve always tried to hire students that are hungry. I’ve I’ve used that word that really want to change the world or or prove themselves or whatever. Uh I don’t want to be motivating students and having to tell them to to work hard or I I don’t want complacent students. I want ones that are hungry that they want to take advantage of this. And and for many people, the time in graduate school is the best time in your life to do breakthrough science. So don’t waste it and and take a really safe project that you know you can do, but rather do something stark and and important. Get out of the comfort zone. Yes. Stay hungry, stay foolish. Yes. Yes. Yeah, that’s true. It seems to be part of the um part of the DNA of this particular sub field of physics because Tobias Kippenberg is also really into triathlon because I remember visiting his group and they were like, “Oh, are you joining the triathlon thing at 5 a.m.” And I have bicycled with Tobias on several occasions. He’s a very good biker. Yes. And you were biking for 27 days or something like that? Well, there’s a ride from uh Tour Divide from from BA to Mexico along the Continental Divide and it was it was I mean in all the rest of my career before that and most of us don’t have the opportunity to take a month off and just you know go go have fun. So I was I was grateful that I could um and I got to see a lot of very virgin territory that I would no most people never get to see along the continental divide. So, but it was very hard um because it was 100 miles a day. Wow. On a mountain bike with all of your stuff and food and everything along rocky trails and and so cuz I read a quote from that I think it was in the Independent that oh it was 100 miles per day. It was actually harder than I thought it would be. And I was like what are you talking about? This sounds hard. Well, there were stretches where it would be literally like 5 in the afternoon. We’d only gone 30 miles cuz it was just, you know, part of it is the trails are just very rocky and you can’t you’re not biking. You’re you’re pushing your bike through, you know. Yeah. Or or a lot of mud or whatever. Um so yeah, it pushed me. It was very hard. Um but uh I was so happy to have done it. I’ll never do it again. But as as the pioneer in in fatonics, I mean we have been discussing like different directions what is going on in the field. uh what would you recommend to people who are would like to make contribution in the field of photonics what what kind of problems should they target well there’s it depends which areas so I mean in in data center sorts of applications just getting higher and higher capacity so faster modulators faster detectors more efficient devices because power getting out is an issue lower voltages so again you know a lot of modulators require three four volts Well, you know, 100 gigahertz CMOS driver chip is not going to is only going to supply you a volt or less. And so lower V pies matter and so things like thin film lithium out work very well for that. Um so those are all good problems. I think for optical clocks and things we need, you know, narrower line with lasers and and a wider variety of wavelengths. Often you need very disparit wavelengths, but again heterogeneous integration works well for that. You know, you might have a a 440 nanometer gallium nitride laser next to an gallium arsenide 860 laser because you need these widely different wavelengths for the optical clock say and uh what’s the most I think a really exciting one is is the whole quantum field right so whether it’s photonic quantum or uh you know ion or neutralbased quantum computing you typically need to to get you know thousands thousands, hundreds of thousands, a million cubits, you’re going to need a lot of Yeah, it’s laser based technology and and so again, you need all these same things. You need very high yield, very good uniformity and uh and so I mean, I think that just we’ve just scratched the surface there and I think I think we’ll make tremendous progress there on the backs of this whole camos technology. What’s the most materials you’ve ever mixed on one? Um, as soon do Joel Guo did three. So, he did uh covered most of the transparent band of optical fibers. So, you know, C to L band was one epimeaterial. O band was another epimeaterial and he had Eband in there as well. Eband isn’t used very much yet, but um that’s the most we’ve done. And so how do you protect then the existing things if you bond again new pieces? Well, one thing is we bond them all at the same time. So you take all three and you you flip it or maybe more than three typically you know they’ll have say 3x3 arrays of it. And so um nine pieces you’d bond you etch off the substrate and then fortunately well if it’s all the same material system almost all the steps can be the same. So if it’s in in gallium arsenide phosphide system etch steps and the contacts are all the same for whatever wavelength material you have. Um if you’re combining gallium marsenide with uh india phosphide based then that’s harder but to a large extent you can use the same chlorine based etching systems and the same we typically use gold contacts but indeed you know Intel and and tower and amal have developed non gold contact that be what do they use instead or t nitride and a bunch of other things. Yeah, there are other. I mean, that still looks like gold. Exactly. Fake gold. Well, titanium’s okay. Nitrogen’s okay. But again, you know, all of our heterogeneous processing in the back end. We’re in the dirty end of the seaoss. So, you know, there’s already copper and and you know, indium and gallium and arsenic and phosphorus are all dopens. So, they’re already there. And so, yeah, um it it it is something you can integrate into the into the back end of the process. How difficult is to introduce this new processes because foundaries are typically very conservative and stick with one with their recipes or even in the clean room like at the clean room they’ll be like no oh you just I’m not going to say it well for UCSB our clean room was focused on compound semiconductors so we had no issues there but when we worked with Intel they first put it in a 6-in fab and uh it was in Israel and brought the process up and then they moved it to an 8inch fab in southern you a 200 millimeter fab in southern Israel and then they moved it to New Mexico for a a 300 millimeter fab and again in each case you know it’s not the most modern fab it’s not doing 2 nanometer lithography but you know say 45 nanometer lithography um and so you know as the premium applications computers and things like that move to more modern ones these old fabs are looking for a reason to exist and and uh so this one exists along with you a lot of automotive electronics and and uh indeed they were very careful to ensure that as they introduced machine by machine that they didn’t you know cause any problems with the rest of the process line because to this day you know like in Intel’s case most of the foundry is running you know electronics not silicon photonics so we can’t take it down and uh so we did that first with Intel we did it next with tower um company I started Orion which is now open light got bought by Juniper um runs 35 in towers 200 millimeter fabrication process and and now there I mean they have a very big silicutonics business. It’s really consuming which is very exciting. But for on cheap interconnects you you do need access to the state-of-the-art electronic foundaries. Uh so again that’s sort of a religious thing that you’re getting into. So some groups like you know labs as is an example use a global foundry process where the electronics and the optics are all in the same thing and so it’s you know in fact Rajie Ram here was famous for developing this zero change approach right so we use a conventional CMOS process and and do it all together personally I’ve never been in favor of that because as clock speeds keep going up you need very advanced electronics uh to make a you know 100 gigahertz amplify ifier or even now 200 or 400 GHz amplifiers has to be very advanced electronics and that’s very expensive and it’s okay for electronics because electrons are small and transistors are small so you can get you know a billion devices in in a fairly small area but the problem with light is the wavelength is a micron and so the size of the devices is large and so if you take a pick and you did it in this very expensive process it’s a very expensive So if you run it in a 45 nanometer process, literally the cost per square millimeter is 100th. And so now it’s it’s 100 times cheaper to have the electronics on one chip and the photonics on a different ship and just just copper bumps bond them together. That’s that happens all the time. Divide and conquer. Yes. So I think that’s a better way to go and I think that’s what will be the dominant approach. What advice would you give young researchers who want to work on ambitious real world systems without losing depth or becoming purely managerial because a lot of uh applied physics groups are now set up like tiny startups that so that you do the managing more or less. I mean, I think the time when you’re in graduate school is a unique time to actually do fundamental breakthrough research yourself and all of us as we get older do move more into managerial positions which allows you to have more impact by supervising more people but less personal satisfaction that you did something right but I think it’s fine to take you know solace in in other people doing well right and so if you help someone be successful That’s kind of the best best opport that’s the best thing. So, is there something else we didn’t cover in this podcast? We covered everything, man. You’re multitasking. I’m very impressed with the amount of research you did on this. I mean, like reading the independent. I mean, I don’t know. Oh, but I also really like reading this these things. I I mean, we do this for fun. We’re learning, too. Yeah.