Optics for the Cloud PhD Event 2020 - Day 1
Okay, well let me kick off so uh i'm at rostron. I'm the deputy lab director of microsoft, research in cambridge, and it's it's my pleasure today, to welcome you to this optics for the cloud, phd, event. Um. It was about three years ago. Or four years ago even i think when we started talking about this internally. And we said uh, we're extremely interested, in how optics could revolutionize. The cloud. For, optical networking, storage, and compute. And we were very keen to engage. With the larger academic, community. And we pulled this together, as a mechanism, this optic to the cloud research alliance. As a mechanism. To allow us to interact. And to sort of be involved with the community. Initially. Sort of very much in the uk and then in europe, and more recently, i know we have people involved, uh who are also in ucsb. And places like this so we've really expanded, up. Um so it gives me a great pleasure to welcome you. This is our phd, event, um. This is the first, virtual, event, for the optics of the cloud, uh research alliance. It's also i think for microsoft, research, one of the first virtual events that we've done so we're all learning i hope that you enjoy what we have, today. Um. Can i go to the next slide please. So, uh you know putting a day together like this whether it's in person, or virtually. A lot of work. Uh i'd really like to thank all of the sort of the people involved. Uh, on the sort of the scientific, chair side we have james and francesca. I think francesca. You're doing, uh today, and james may be doing tomorrow. Um. We have a number of speakers, also who have also been heavily involved from the microsoft, side and putting all this together. Hitesh, pablo, and paulo. Uh, we have the team that have done all the logistics, and organization. Uh scarlett, claire and lynn and i'd like to thank you all, and also ben has also been instrumental, in pulling all of this together, getting the program, ready, and. Making sure that we have a, good set of, talks. For these uh next, sort of two afternoons. Uh, can i go to the next slide. So today's, agenda. Uh starts off with me welcoming, you which i've almost completed. Uh hitesh. Will be giving you a sort of a whirlwind, tour of optics to the cloud. I've seen the number of slides it's got it certainly will be, whirlwind, as you go through them. And then we're going to go into a number of talks. Um. From the phd, students now, all of these phd, students are special because, they're all funded. With or by microsoft. And. We have people internally, inside microsoft. Who work very closely with their academic. Uh, mentors, and advisors. In order to help make the work, uh and help to input on the work which, uh to me is a great, you know great thing, i, often say that microsoft, research isn't a funded body we're not like the you know like the epsrc.
In The uk, or the nsf, we don't just write checks. And that's it we like to actually be involved and we like to partner, and so, it's great to see these talks happening and people from both sides and universities, as well as from microsoft. Are, involved, in them. So, that really brings me towards the end of what i want to say can i go to the next slide please. Uh, it sort of leaves false to me as well, to, uh you know to sort of give you some of the housekeeping. Uh, some of the thinking for how we would like today to operate. Um. I can tell you having done other things online, that, if everyone is completely, silent. It's not it's not great, um. We have chat windows, so you can ask questions, in the chat windows, and there are not just the speaker but there are others online. Who are going to be able to answer, questions. And i would encourage those from microsoft, and those from the universities. If you see questions coming in where you think you can help answer them for the speakers. You should dive in and, provide some of the answers you know get a little bit of sort of almost, interactive. Uh discussion. Happening, in these, in the talk associated. With the event. Uh if you'd like to ask questions. Uh i think certainly, inside the microsoft, ones they're all going to be live. So you're able to uh raise your hand. Uh that will attract. The attention of the session chair, today is francesca. Uh or you can send a message in the meeting, chat. Um. Some of the talks are recorded, so obviously it's not gonna be possible to stop them in the middle, to answer questions. For the live ones if you've got something that you think is really important to be honest you should just ask it and put your hand up and make it interactive. Um. We should also you know i think for these talks it's a great opportunity to collect, feedback. If you have thoughts or comments, that you would like to share please feel free, to share them with the speakers, either through email, or through. Going through us. After the event. Can i ask everyone, to stay muted, you know i think we've all noticed this tendency at night some of these online meetings, for, uh you know background, noise, and whatnot to come through and obviously if you're trying to talk and there's a lot of background, buzzing and hissing coming through it's a lot of distraction. So let's please, turn those off of course unless you want to speak. And i i would encourage people if they can ask questions. To actually switch the camera on i think it makes a huge difference in these online meetings. If people have their cameras, on when they're interacting. Um. You know we we have to raise the hand again which i've just mentioned. And, um, you know claire morgan. Who's, online, and we've got an email address for her here, if you're having technical problems, please feel free to send her an email, and we'll do what we can to sort it out. Obviously with these sort of things it's not always possible, to sort out. The problems. Remotely. I've discovered that my house bandwidth, is a real problem so i'm looking forward to some of the optical networking. Talks to understand what we're going to do to sort, the bandwidth to my house out, but anyway. Um. You know so. Let me just conclude this bit by saying you know please, try to be interactive, don't just think of this as i'm going to sit here for a few hours and just listen. Try to think well you know answer, ask questions, be involved. That would be the basic message. So i think that's my last slide. Uh scarlett can you go to the next slide is there another slide. No i'm gonna head over to hitachi, now, okay so, i'll hand over to justin so here we go so uh hitesh is our first speaker. He's going to, um, give us a talk, uh a whirlwind, tour of optics of the cloud here at microsoft. And uh hitachi has been with us for many years, i think coming up to 12 years or 14 years is it hitache, it's it's, not 129. Is it. Is it barely 10 you're a youngster. Um. And, uh, you know it's one of the key people one of the key leads, on all of our optics the cloud work. So i'd like to hand over to him and uh. The floor is yours. Ah thank you ant uh, scarlett, is giving me control, i'm just gonna check can i do a quick audio check and or francesca. Could you just say are you a perfect, excellent thank you, so, um. It's great to see so many people here thanks and thanks everyone for your time. And, what i want to do over the next 15 minutes as i mentioned. Was. Give you a taste of some of the work happening, here at microsoft, research. As we push these optical technologies, for the cloud. And. I mentioned this in the past. Give me a second it will just take me a second to get control. You should have control a heater. Yeah. It just takes, 10, seconds. That was. So interesting. Yeah it was. Um. It was a very nice talkie tesh thank you very much i knew that. Well as ant mentioned it was a whirlwind, tour.
Apologies, For this we're learning on the fly. Just give us a sec. Please. Thanks color. Let's try this again. So again this takes 10 seconds, but i'll get started in the meanwhile hopefully i get control, of being able to move the slides so. Um, and something that i've mentioned in the past is that we're really, worried about the fact, that almost, all technologies. Underlying, our cloud infrastructure. Seem to be approaching, the you know tail end of their exponential, growth phase. This applies to. Compute. It applies to networking. And, if the slide moved, you'd see that it would apply to storage, too. And so, what we are really looking for is. Can we leverage, bleeding edge optical technologies. That could lead to new growth curves in the post moore's law era. And you'll see this as a common, theme, as. Across the projects as i talk about them. So, i'll start on the storage print. And, one of our big projects, in this space. Is silica. So, this is a collaboration. With azure storage. And what we are trying to do is see if we can disrupt. The archival, storage space. The motivation, is simple. The amount of data in the cloud, just keeps growing and growing, and most of this data, is archival, in nature. I'm sure, all of us have, thousands, of photographs, of cats, and dogs, and babies. Personal, records, and financial, information. Astronomical, data, weather data. And today. All this is stored in archival, storage. Spools of tape. However. This technology, is finding it hard to continue, scaling. Not to mention, even with these technologies. Data needs to be copied, every five years, in order to avoid bit rot, which means, high cost. And high complexity. So. The aim in silica, is, can we use, glass, or silica, as a storage media, in order to reduce, or eliminate, this gap. Now. The underlying, technology, that enables this work is femtosecond, lasers. And the basic concept. Was developed at the university of southampton.
You Shine. On a piece of glass. And its focal, point. We get a nano grating or a voxel. By changing the intensity of the laser, and by changing its polarization. We can change the size, and the orientation. Of the voxel. Hence allowing us to put multiple, bits on each voxel. And can create, layers of these voxels. Essentially, getting a three-dimensional. High-density, storage medium. The other advantage of this technology, is that allows us to separate. The right, technology. From the read heads. So the reading in this case, is done using computer, controlled microscopy. We shine a polarized, speed of light, that can be moved, along the x y axis. And we can change the focal plane. And this received, light is then decoded, using machine learning, which in turn, allows us to achieve, higher density. As opposed to traditional techniques. Now. One of the key advantages, of using glass as a storage media. Is its. Amazing durability. Thousands of years. So, we've boiled our samples. We've magnetized, them, we have scrubbed, them, we have microwaved, them and, we have baked them. And in all cases. Not only did the sample survive. But we were, always, able to recover, the entire dataset. Now this is something that got us really excited. And it led to external collaborations. For example, we have a big collaboration, with warner brothers. And as you can imagine. They have, lots of high value, archival, data, store. Tons of movies, of cultural. Uh importance. Which today, are stored in these. Spools, of archival. Tape, in temperature, controlled. Expensive, data warehouses. So, we work with them to write the original, 1978. Superman, movie, on a piece of glass. Which was actually showcased. In satya's, keynote at the end of last year, so our ceo. Uh it's a small piece of glass, um, 75, millimeters. By 75, millimeters, so, like a coaster. We wrote around, 76. Gigabytes, of data. Plus, some, redundancy, information, across, 74. Layers in this class. And this was at the end of last year. Our density, information, our density, status, has improved significantly. Since. Now this is obviously. An achievement that we are proud of it led to advertisement, for the company. It. Allowed us to have this collaboration. With partners. Who have high value archival, data, and who are willing to work with an experimental. Futuristic, technology. So there's a meta point i want to make here, regarding the journey, from. The optics, to the cloud. So in this case the building block technology. Is femtosecond. Lasers. Amazing, piece of technology. Are developed in a physics lab over a decade. However, it's still a long long way, from what i'd call, an, end-to-end, storage system.
Because, A storage system like that, needs to have a full-blown. Right pipeline. It needs to have a glass library. It needs to have a read pipeline. And this entire, system. Needs to be designed. With, high throughput. High capacity. Fault tolerance. And of course it needs to be coupled, with a cloud, storage software stack. And this is exactly what the team has been working on over the past two three years. And i think. The main reason we've been able to answer this feasibility. Question the affirmative. Is because of the cross-layer, nature of the team. So we have. Hardware architects. Software engineers. Physicists. And chemical engineers. Working in the same team in the same labs. And this is the point that pablo will make in his talk tomorrow. How we are now looking at ideas, from the networking, domain, that could improve, our read and decoding, technologies, even further. Of course not all the technical. Questions have been answered. We've gone from this. Basic feasibility, mode can we do it, to the. Should we do it, getting this technology, ready for prime time, which has both, technical, questions, to be answered but also strategy questions to be answered. So, the next thing i want to talk about, um is in the networking space and this is a project called series. And the idea or the motivation, here is, simpler. Historically. We've relied on this exponential. Growth, of the free scaling of electrical, packet switches which has been great, but things, are looking gloomy, and lumiere. And so the question we are asking is. Can we use, optics, to, generate, a new switching, growth. Now, there are many switching technologies, optical switching technologies, out there um some of them have been explored, by people in the audience for many years. And actually this is a question. That the community, has been asking as a holy grail for the past 15 20 years. The particular technology, we pursued is actually very simple, uh it's a grading, a piece of class with etchings. Or an ewgr. So if i want to speak to francesca, i'll send my data on the green wavelength. And if i want to speak to scarlet, i'll send my data on the red waving. So we can use this as a switching technology, in our data centers. By coupling it with the tunable laser. However, in order to support. Bursty cloud workloads. We need to be able to switch, in a very fast fashion. Now a second granularity. Which means, a tunable laser, that can tune its wavelength, in a very fast fashion. Since this is, hard to do with. Existing, off-the-shelf, tunable, lasers. We took the dive in the fabrication, space. And we designed, our own custom, pick. That can tune over the entire, optical, c band, in less than a nanosecond. And we have, multiple. Implementations. In, multiple, variants of this design. Now. From a chip perspective. From a technical perspective. This is something that, you know i think is very neat and i understand i'm biased. Uh it was a great learning curve to go through this fabrication, exercise. However, i want to go back to my meta point. This is a very useful, building block. But it's still. Miles away from an, end-to-end. System, a cloud network, offering. Nanosecond, switching, at end-to-end, granularity. And in order to get there, we have to solve. Several. Hard, systems, challenges. At different, layer of the cloud network stack. So just to give you an example. The awg, as mentioned, and the. Tunable lasers, have around 100 payments. So we can use this technology, easily to connect 100, servers. But in the data center, we need to connect, 100, 000 cells. So how do we scale. So. We had to solve, all these questions. In order to build a prototype. That can offer. Nanosecond, granularity, switching from an application, perspective. And paulie will describe, in this talk tomorrow, as to how we got there. And i think the secret sauce. Was the fact that we didn't go off these problems. In a layer by layer fashion, with point solutions. We actually, came up with a cross-layer, design. Which means the solutions. Are practical. And we think deployable. However there's still a long way to go you know we moved from, can we do it phase. Should we do it first. We understand the design space, we understand, the trade-offs, that different solutions, bring to the table. And now we are trying to identify. The, best cloud scenarios. Where this technology, makes more sense. So that was the. Work in the networking, space, and like storage, this is stuff that has been happening over the past, three, maybe even four years. So the next thing i want to talk about. Is something completely, fresh of the press um compute.
The Motivation, again is simple. Moore's law is slowing down, and can we use. Optics, for compute exploration. Now. I'm sure, most of you have thought about how optics can you be used for different, computational, primitives. Just a simple example, if i were to take an image in an optical domain, and i shine it through a lens, on the focal plane i get the image in the fourier domain. Which can be used for things like pattern matching. Similarly. If we wanted to do vector by matrix multiplication. We can represent, the vector, by an array of pixels. We can have our matrix. On a spatial, light modulator. And on the output side, on an array of cameras, or photo detectors. We get the dot product, of the vector, and the matrix. Pretty well known concepts. However there are three, things i want to point out here. First. This is analog, computation. So the precision, is limited, as compared, to a digital counterpart. Second. The competition, here is, free. Once you're in the optical domain. But it does require. Us to go, from the digital twitter to the optical domain and then back. And finally, i want to reiterate, that we are not talking about general purpose computation. We are talking about, acceleration. Of specific, computational. Primitives. So we could imagine, using these. As an optical, coprocessor. Or an optical accelerator. So imagine. Using this optics. For accelerating, ai inference workloads. In this case we have an image of a car. Which is. Processed, by a digital engine, which is implementing. Some parts, of a new neural network. It's converted. To the analog domain and then we invoke, our optical, accelerator. To do our vector to matrix multiplication. And finally, we go back to the digital domain, with the answer saying well this is very likely a car. Now, this is an interesting pipeline. But if you notice. Every time we invoke, the, optical, accelerator. We need to take the tax, of the digital to analog, and analog, to digital conversion. And by the time we account for the power and the cost and the complexity, of those conversions. Most of the benefits, here seem to go away. So even though there's a lot of hype regarding, the use of optics. For accelerating, ai influence. When you look into the gory details. The benefits, are not that spectacular. However. We don't think this is a dead end. There are, other computational, models, where the pipeline, is different. For example, if you look at. Solvers, for combinatorial. Optimization, problems, which are prevalent, in when you're doing, finance, portfolio. Optimization. Or you're doing molecular, drug discovery. So in this case we have an optimization, problem. It's converted, into the analog domain. And we invoke, the optical, accelerator. Hundreds, of thousands. Or tens of thousands of time. And we iterate, to the optimal solution. And once we get close enough, we go back to the digital domain. So this is an. Application, scenario. Where optics can really shine. Because. The overhead. Of the, digital, analog, and analog to digital conversion. Is amortized. Across many competitions. So this is a, a particular, computational, model that we are excited, about and that we are pushing on. However, its early days. We are still in the, can we do it phase. And hopefully we can come back, in a year or two, and answer that question in the alternative. So to step back and summarize. Uh. We are generally, worried about the fact, that compute.
Networking. And storage. Seem to be approaching a platform. And we've taken this bet. On. Developing. Optical technologies, that could lead to new growth curves. And we really believe. That this, cross layer interaction, and cross layer techniques. Are going to be the key to success. On this journey for us. So thank you for your time hopefully that was useful. Uh happy to take questions. Verbally, on the chat window over email. I'm here for the next couple of days. Hopefully everyone has a good workshop. Thank. You. I don't know how we are doing on time do we have time for a few questions, if possible. Yes i think we would have time for, a couple of questions, if there. Are. Any brave, question, speakers, i see somebody. Unmuted. Yeah, so there is thomas, uh. Gerard. Please. Thank you john. Hey dash, thanks for your talk. Um. You, spoke about, uh whether the, obstacle switching, solution. Is the, if you should do it, for the cloud data center. That's the the work stage you're at now and i'm wondering what other kind of application, spaces, it might be suitable for and where you think it has a lot of opportunity. Yeah. Yeah so this is a point, excellent question tom so i'll give you the sort of the 30 second answer and i believe paulo will later today will elaborate, on this, so there are different, kinds of workloads, where, optic switchings. May or may not shine, so just to give you a concrete example. If you think about how we think about, servers, and data centers. Uh they have a bandwidth, requirement of, 50 gigs 100 gigs. So if you are trying to connect. Optical switching, for, those sort of scenarios. It provides benefits, but it's not, as compelling. Because the electrical switches are doing a pretty reasonable job. On the other hand there are many many scenarios. Like um. Ai scenarios. Where these digital ai chips, have a bandwidth, of, hundreds of kicks, actually, the latest, nvidia gpus, have around, uh 2.4. Terabytes. When you start thinking about the kind of switching requirements. Both from a power and a cost perspective. For. Let's say ai, training and ai inference. That's where optical switching may really shine, so that's what sort of the meta context, behind that. Comment, about scenarios. You really need to think about, the bandwidth. The latency. And, other requirements. Um and of course things like deployability. In order to identify, where this technology, may or may not shine and i think this is the point that paulo will, make in his talk i saw him so maybe he can comment on this. You know i just wanted to add that latency is also the other dimension, and that is very important. Which, for traditional, traffic where is vm to vm, most of the latency, goes into the stack. So reducing. You know from you know microseconds. To, nanosecond, doesn't really matter, but once you start going for some of the more, aggressive, workloads, like resource disaggregation. Or. Memory disaggregation. Then latency, start you know becoming like a primary bottleneck. And so the ability effect of having you know a solution. That can, give you ultra low latency, becomes very compelling. Thanks guys. Other questions, so pashmina, had a comment that yes uh so pashmina, says that uh the kind of, comments being made regarding. Optical, computation, would be good for any iterative, algorithms, as long as one can evaluate, the goodness in the optical domain not just combinatory. Optimization. But you know that's a spot on comment yes it was a, actually the general comment there was about, if your computational. Pipeline, looks like this that you can. Remain in the optical domain as long as possible and hence amortize, the overhead of the d2a, and the a to d, and that, those are the scenarios. And yes we should have internal discussions, and external discussions, on, the kind, of application, domains. Where this computation, model holds. Fantastic. I don't see any further, question, for ritesh. But please feel free to keep using, the meeting, chat for, any further, question, that you may have, to. For him. But, without, further, delay i would also like now to introduce. The. Next, uh. Four or so exciting. Talks from uh. Phd, students, who are collaborating. With us. And, we will then conclude, the day with the last, microsoft. Talk. Um, so without, further, delay i would like, to introduce, you, uh the first speaker. Arslan.
Raja. Arslan. Is a member, of the, laboratory. Of photonics, and quantum measurement, group at, epfl. Under the. Supervision. Of, professor. Tobias. Gibenberg. He works, on photonics. Integrated. Multi-wavelength. Sources, for data centers. Actually, he joined, arsenal, and joined us for an internship. With us last summer, working, on project. Series. Under the supervision. Of, sophie, lounge. And today his talk is about. Sub nanosecond. Optical, switching. Using, multi-wavelength. Sources, based on solid on microcomb. So let's start. His, pre-recorded. Talk. And. This will be followed, by question, and answer where arslan, will be there to answer, any question, that you may have, lived. So thank. You. Good afternoon everyone, my name is ashland sadie raja and i'm a phd student in the lab of professor tubby skippenberg, at epfm. This work is done in a collaboration, with optics for the cloud team at microsoft, research, cambridge. Today i will talk about sub nanosecond. Optical, circuit switching, using, solid on microcomb. Solar on microcom, is a multi wavelength source, or until natively, it can be considered as a mission of multiple, laser from a non-linear, element which is micro generator, a circular device, shown on this photonic, chip. I want to start my presentation, by giving the motivation, of our work. This figure shows the, multi-tier, architecture, of current data center, in which, servers are interconnected. With the help of electronic, switches, and optical, fiber. The black line represent, optical, fiber, in this figure. The data is converted, from electrical, domain, into optical domain and from optical domain into electrical, domain using optical transceivers. The conversion, process, add some overhead, and also these devices, are power hungry. Due to the massive flow of data within the data center, and also to the external, world, causes a lot of issues, related to the power, and, scaling. The one of the main reason behind, all these issue is, is following, all the electronic, component, used within the data center are reaching the, performance. Limit. In order to solve all these issue wavelength based circuit switching, is proposed as a key enabling technology.
This Particular method is used by our collaborator, at microsoft, research, cambridge. The basic idea behind this method is following. The interconnection, between two, server, is only established, using a predefined, color of the light for example, if this server, want to talk to this last server, they can only establish, a link or communicate, data, using, green color of the light. Some, advantages, of this particular. Technologies. Are following. There is no need of optical to electrical, and electrical, to optical conversion. Which mean. There will be less number of optical transceiver, used in the data center and hence more power efficient data centers. The interconnection, between server within the data center are more efficient as compared to the conventional, data center, this particular property is more important for certain applications, such as, cloud computing. The port scalability. Is not an issue as most of the technology. Our method, can be relies on a photonic, chip based, platform. One of the key component, required for optical circuit switching is a multi-wavelength. Source, a multi-wavelength. Source can emit many color of the light. A straightforward, and ideal solution, is to use tunable, laser. A tunable, laser can emit many color of the light, either by changing the temperature, or current applied to this particular. Laser. It has been demonstrated, that this particular laser cannot. Tune faster than 5 nanosecond. And in order to, efficiently. Utilize, the different resources, within the data center it is important to switch at sub nanosecond, time scale. Later it was proposed that an additional, element, known as wavelength selector can be used. To select different color of the light coming out of multi-wavelength. Source. This particular technology, has enabled. Sub-nanosecond. Time scale. Circuit switching. In order to understand this thing in more depth let's consider we have a three laser diode, red, yellow, and blue and they are continuously. Emitting the. Different color of the line, in order to select one of these color, we will, use, a optical, switching element, semiconductor. Optical amplifier. If we want to transmit, the red. Color we will apply a control signal on the this way. By applying the, control signal on first password so we'll see, that this red color is transmitted. Now if we want to switch from red to yellow we will apply a control signal on the next soa. So by applying. A control signal on the second soa, we. Start transmitting, the yellow color of the light. Similarly if we want to transmit the blue we will apply a controls. Apply control signal on the third soa. So by applying the control signal on the soa we can select. One of the color emitting, from the laser diode. This particular method has already enabled. Sub nanosecond. Time scale circuit switching. For example now if we want to increase the, number of channel. If we want to increase the number of channel. Then, a simple solution is to use, array of the laser. For example if we want to operate with 64, different colors, then we need. 64, independent, laser diode. As the emission color, depends on the current of the. On the of the laser diode, it is very important, to stabilize, this current sources. And also. Emission color depends on the temperature, of laser diode.
And It's also important to stabilize. Thermally stabilize, this laser diode. So this bring an additional, complexity, of. Stabilization. And also the power consumption, so an array of laser is not an optimum, solution, for the, increasing, the, number of color, the number of wavelength, channels. One of the best and optimum solution, is to use solid on microcomb. As a multi-wavelength. Source, this particular, multi-wavelength, source can be generated, from a non-linear, micro regenerator. By coupling a single frequency, laser diode inside this microresonator. Due to some non-linear, frequency, conversion, we see many new frequency, components, coming out of this micro resonator. This multi wavelength source provides some key advantages, over the area of the laser which are following. This multi wavelength source can be generated, from a single, chip. Also this. Multi wavelength source provide more than 100 comb lines. And what i meant from the easy wavelength, alignment, is following. Just by aligning one of this comb channel we can align all the other comp channel, similarly. If we stabilize, one of this comb channel. We can stabilize, the all other comp channel. As compared to the area of the laser where we need to stabilize, each color so this provides some additional. Benefit, over the. Array of the, laser. In this slide i want to give an overview, of optical frequency. Mainly the generation, of solid on micro combs. We start with the single frequency, laser diode, we couple this, laser diode inside this micro generator. With the help of bus waveguide. This micro resonator, has a high quality factor mean it has very low. Propagation, losses, and it also, it has. Car non-linearity. At a certain threshold, power and by adjusting, the laser properly, we see, two new frequency, component. With the help of degenerate, forward mixing. If we increase the power further and adjust the laser properly, we see, additional, frequency, component, that are generated, with the help of non-degenerate. Forward mixing. If we increase the power further, and adjust the laser properly, we see a comb-like, structure coming out of this micro generator. As i mentioned in my previous slide, the spacing between this comb line is. Is a. Constant. A typical comb structure, is shown in this figure, where the spacing, between two comb line is constant. And this optical frequency. Can be considered, as a frequency, ruler.
As We know that we use. This normal ruler to measure the distance, this frequency, comb, can be considered, as a frequency, ruler to measure the unknown frequencies. So this particular tool has been used to understand, the fundamental, physics and. It has been also implemented, in many technical, applications. One of the very important, uh, frequency, comp state that i want to mention here is dissipative, causality. If we can properly, design the dispersion, and non-linearity. Of. Of micro generator we can excite, this pulse-like, structure. That will maintain, its shape, inside the microgenerator. While propagating. In time domain is correspond, to pulse of train. While in frequency, domain is correspond to optical frequency, comb, with the envelope, of secant hyperbolic, square. One of the example of, optical spectrum, of soliton, is shown here in this figure. The other key advantage, of this particular, state is following, it this particular, state is, intrinsically. Low noise, so you can think about all this comb line as an individual. Coherent. Laser that can be used to transmit, the data. Or they can be used for the different, relevant. Applications. Now i will discuss the, all the experimental. Work we have done in this project. The first part, is solid on generation, so we started with a single frequency, laser diode. Then we, amplified, this. Laser, in order to. Overcome, the coupling losses so this package device has around 15 percent coupling loss which is very low. If we provide enough optical, power, and, adjust the laser properly, we can excite, a single solid tone, this solid tone has a, envelope, of secant hyperbolic, square. 3 db bandwidth of this spectrum is around 40 nanometer. The fsr, of the micro energy, is 100 gigahertz. So. We can use around 16 coherent, carrier for the, optical switching and communication, experiment. We did the, further, amplification, of the single soliton, in order to meet the power requirement, for the experiment. The post amplified, spectrum, is shown here, we have achieved maximum, power up to -4, dbm.
And Most of the channel has optical, signal-to-noise. Ratio around 34, dpm, which is enough for the, experiment. That we want to carry on. After generating the single solid tone we picked only single comb channel if you see this. Spectrum we only picked a single color from the spectrum, using this element awg. And we sent this signal through soa, which is a optical, switching. Element. And then we detected this signal on the photodiode. So if you see the detected, signal. We can clearly see that we were able to turn on and off this signal. At some nanosecond, time scale. We similarly. Tested. More than 25. Channel in c band, and they showed the sub nanosecond. Switching. In the following, experiment, we have picked four different comb channel, from the solution, spectrum. And then we use four different semiconductor. Optical amplifier, to switch these. Com channels simultaneously. We applied a modified, control signal in order to switch this. Comb channel. The signal detected, on the photo diode shows that we can switch for comb channel. Simultaneously. At sub nanosecond, time scale. The maximum. Optical, separation, between four, comp channel that we have shown is 20 nanometer. Just to mention that we were limited by the optical bad pass filter. Not by the soliton, microcoms. In order to show a system level demo a maxender, modulator. Is used to encode the data. In the form of anarch, and pam4, on the simultaneously. Switch, for solid, on, lines. If you see in this graph, a performance. Below, forward error correction threshold, limit. For the data center is achieved for both nrz, and pam4. Data transmission. At a different, received, optical power. This shows that solar and microcomb. Can be. Used, as a potential, multi-wavelength. Source for the future data center. After certain, improvement. All the experiment, that i have shown till now has utilized, these discrete, sos, which are very attractive for the proof of concept, experiment. But in order to demonstrate, the switching and data transmission, for the future. And practical. Application, which is very important to demonstrate. This application, at the chip scale level. So this chip, was designed, by the microsoft, research cameras team. It contained 19 soa, and awg. So this particular photonic, chip provide the advantages, of compactness. And energy, efficiency. So we perform the similar, experiment, that i have shown till now using the discrete component, first we, perform the switching experiment, using chip based, soa, and awg. Along with solid and micro comb so we picked two. Comb channel. And we showed that it is possible to demonstrate, sub nano second, circuit switching using. Chip based soa, awg. Similarly, we perform a data transmission. With the modulation, format, of anarch, and achieve the performance, below forward a regression. Threshold, limit. Due to some technical issues we were not able to demonstrate, the pamphlet data transmission. We are looking into the possible, reasons, such as comp power osnr, and crosstalk, between awg, channels.
In Order to improve the. Performance, of, our system for the future, uh. Experiment. The next thing that we also want to improve is the power efficiency, of the our devices. So we have identified. Several parameter, that we want to improve, the first thing that, as i mentioned. Previously, in one of the slide that the total, coupling efficiency, of this packaged silicon nitride microregulator. Is 15. Which is, very low. Uh the main limitation, is 2 db splicing, loss so we want to reduce the splicing, loss around 0.2, db this will improve our coupling, efficiency, around 30 percent. And increase the power efficiency, of our of our system by two times. The next thing that we want to improve is the dispersion, of the micro generator. By engineering, the waveguide, width and height we can. Engineer, the. Shape of the spectrum, that will be generated, from that particular, micro generator. This will, eliminate, the need of post amplification. And the spectrum, can be directly, used to perform this optical, circuit switching and the data transmission, experiment. And by eliminating, the need of post amplification, we can improve the power efficiency, of our, device. It was recently, demonstrated, that solid and micro com source can be, generated, from a on-ship, laser without requiring. A frequency, stabilized, laser. This fully integrated, solar turn microcom, source can be interfaced, with fully packaged, soa, and awg. Photonic, chip, this will provide, a path toward fully integrated. On-chip, sub-nanosecond. Optical circuit switching and data transmission. Before. Finishing my presentation, i would like to show the different application, of solid on micro comps. In order to show that this solitude microcom, can be used in many different applications. The first application, that i want to highlight is a coherent communication. In which a solar and micro comb can be used to transmit, data at the rate of 55, terabit, per second. Similarly, solid and microcom, can be used to, perform the medical, imaging. Very recently, our group has shown this thing, that solid and micro con can be used, as a parallel, source for frequency, modulated. Lidar. And it, it can be also used as a calibration, for the, spectrometer. To detect the exoplanet. So this shows that solid on micro comps cover a wide range of applications. Before. Concluding, my presentation, i would like to acknowledge, certain people i would like to thank them. Optics for the cloud team at microsoft, especially, sophie, and kai who carried out the, most of the experiment, that i have shown in. In today's presentation.
Similarly, I would like to acknowledge, my group at the epfl. Especially. Maxime. Zin. Anton, and. Junky, and my professor, prof, my professor dubai skippenberg. And, thank you very much if you are if you have any question please feel free to ask thanks again bye. Thank you very much. Arslan, for the very nice, talk. Um, i believe arslan. Is. Here with us to answer, any question, that. We may. Have. Yeah yeah can you hear me yeah yeah. Perfect, yeah, so i think it was, asked, on the, meeting, chat. About. The reasoning. Um. Of uh. Not getting, down, for br, not to go below 10 to the minus nine. And so he answered. Um. Due to the. Length of the data sequence, and. Nicola, calabreta. Was still asking. Which sequence, it was used. Already, answered, yes, any other question. For, arsenal. Yet tom. Uh, gerard. Has a question. What's, the optical, conversion, efficiency. For seed laser, into the micro ring resonator. So the conversion, efficiency, of. I mean from pump to the to the adjacent, comb line that are in three db bandwidth insulator, microphone, is around two percent. So the net power existing, them. Okay so if you take out the pump then i think we have around uh. Four milliwatt, of power coming out of uh. From this micro resonator, so. Coupling, the conversion efficiency is two two percent and the net power of com that are basically coming out is around. Uh four milliwatt. And if you want to know the absolute power of each comb line i think we have like around minus, 20 dbm, so then we did the post amplification. As i mentioned in my presentation. So. The power, for sure is not very high. In the in the comb line so one of the reason i also mentioned in the presentation, the coupling was very low it was around 15 percent so we were not getting too many. Uh. Not getting too many uh, you know, power in our comb line yeah. Okay. Then. Uh. Okay yeah then, george's, service, as, another, question, is asking. What is the definition. Used for switching, time. Rise. Or, cycling. And what was the overshoot, that appears. On the scope. So the rise and fall time i think they it was like uh, from 10 percent to, i mean like a standard definition, from 10 percent to 90 percent of your signal, okay. And the overshoot, that we saw on the chip base experiment, it was i think there was a mismatch. Between the probe that we were using on the on the on this soa, for the chip base that's way so, there was an impedance mismatch that led to this uh. Overshoot, in the switching. Signal. Yes, so then, we have another question, is there any specific, reason, for using, silicon, nitrate. For micro, for the micro ring resonator. So yeah i mean it has uh okay, um. It has very low losses as we know that the. The, the initiation, power really depend on the q factor. First thing it has also relatively, very high carbon linearity, which is required for generating, these micro combs. And, also, i mean we can, fabricate, this thing at the chip scale level so this is like a three normal, like a top level reason i can tell you here why we use the silicon nitride. Perfect. Then, just to be on time i don't see any further, question, on the chat but feel free to, approach, arslan, by email, if you have any further, questions.
I Will then like to move on to the, next speaker. Who is, wenjin, kyan. Wenjing. Is a, member, of the photonics, integrated, group at, tu endoven. And, the supervision. Of, professor. Javier, legend. Kaishi, is a professor. He's a. Microsoft. Supervisor. She's, working. On a hybrid, integration. Photonic, circuit, system, based, on indium, phosphite. And, silicon, nitrate. And their corresponding. Integration. Which is one of the greatest, barrier, to deploy, optical, switching, technologies. And today's, her talk is about, cost-effective. Assembly. For, hybrid, integrated, photonics, switches. And. Let's start. Pre-recorded. Talk, which will then be followed, by, live. Qaa. Where, ranjin, will, be able to answer. Thank you. Hello, everyone, this is wang jing. Thanks for joining my presentation. Today, i will. Talk about. Cost, effective, assembly, for hybrid. Integrated. Photonic, switches. Let me show the content, of this talk. First, i will briefly, introduce, the field of photonic, integration. And give you a sketching, of our project. Then i will spend, most of my time on our proposed, assembly, concept. And three designs, of photonic, integrated, circuits. As demonstrators. In this assembly, route. Say, a photonic, switch. A multiplexer. And transmitter. There are three key elements, in data centers. Range from optical, source. Over wdm. To optical networking. Of course, since time is limited. I will not go into detail about each design. I will mainly focus on the photonic, switch. Finally, i will give a short summary. Let's start, with photonic, integration. What comes to your mind when you think of it. Basically, photonic, integration. Is a technology. Of combining, many optical, device, components. In a single circuit. Interconnected. By waveguides. We call it photonic, integrated. Circuit. Or pig. It's enabling, lots of applications. For instance. Telecommunications. Healthcare. And data center networks. This widening, range of applications. Can be implemented. By one generic, integration. Technology. Here are some examples, of pigs, based on indian phosphide, platform. As you may know indian phosphide, is a direct, band gap material. And could generate, light. It can enable, various, integrated, lasers, such as ocd, lasers. Wdm, transceivers. And lidar. It can also benefit, biosensing. And recent quantum, key distribution. With these kinds of pegs, we can achieve great performance. Such as reduced, footprint. Improved, stability. And reduced, energy consumption. It's also good to know how the pegs presented, here are roughly five millimeters. Long. Naturally. This benefits, can be extended, to data centers. Microsoft. Is actually using the same platform. To investigate. Data center applications. For example, fast optical switching. Imaging, a data center. There are hundreds, of thousands, of servers, and rpo transceivers. In such a complex. Optical transmission, system. Optical, interconnects. Between, fiber, and photonic, circuits. Are happening, very frequently. So efficient. Fiber to tip coupling, is crucial. However. While facing a huge, size mismatch. Between, a fiber core. And an optical waveguide, on trip. It can cause considerable. Transmission. Loss. And limit the scaling, up of capacity, density. To date, various, methods, have been proposed. For example, using an. Interposer. Between fiber ray and photonic, dye to find in. Or using, a fiber stop. To match with wave booth array, with polymer, buffer, integrated. On chip. Wirefly, speaking. Current, solutions, can either support, many connections. But in a tight alignment, tolerance. Or. It can. Give a relaxed, tolerance. But limited, connections. Another, concern, is photonic. Assembly, and packaging. The gold box packaging, make it even more difficult. For a large scale, fiber recovery. Not to mention the gold box packaging, is quite expensive. All these challenges, drive our work, to find a solution. That can enable, many connections. With relaxed. Tolerance. In a cost, effective, way. Now let me introduce, what we are doing. We are investigating. An assembly, method, for hybrid, picks. Especially. For phosonic, switches. By leveraging, two technologies. Indian phosphite. And traplex, technology. Few more words about triplex. Is based on alternating. Silicon nitride, and silicon, dioxide. Films. To take our, assembly. Method, a bit further. As the picture, shows. A triplex, die is utilized. As a kind of photonic, vcv. On which indium phosphide, paradigm, will be electric, bonded.
Particularly. Flexible. And free-hunging. Wave guide and traplex. Are connected, with, a fiber ray, and will be aligned, onto, indian false light. You may be interested. In the, flexible. Wave guides on traplex. We will see how it works. Let's move to the assembly. Concept. As for the flexible. Waveguides. They are implemented. By silicon, nitride, based, treblex, technology. Because, of fingers. Each finger, has a waveguide, inset. As shown, in the schematic. Build. Tle, has developed. This method, to enable, optical, interfacing. You can regard, the fingers, as spot, size converters. Between, a fiber ray and chip. You can quickly, say. This is, a finger. Array. One side connected, with, a five array. And the end of fingers. Will be, aligned. To, a chip. This approach, can support, up to 40, connections. Per millimeter. As shown in the picture, here, this is an example, of. 64. Fingers. With a page of only 25. Microns. You may ask, how can the fingers. Breach a fiber ray and chip, we will say it later. In this slide. You can clearly, see how the triplex, fingers. Are aligned to the final shaped landing zones on ending full side. Especially. In these two pictures. Picture, c. And picture d. This is the alignment, procedure. Of an array of four fingers. In the upper one, before, alignment. The four fingers, are moved to somewhere, above the tube. And in the lower one. They are landed. In the funnels, on indian fall slide. And then, aligned. To the waveguides. Here's, an another, example. Of more than 30, fingers. Alignment. In the hair level. When the indian, phosphate, dye is being phelped, or bonded. Triplex, fingers. Can be passively, aligned, into the, dedicated. Funnels, on indium phosphide. Without, optical, signal monitoring. So, this is the general idea of our assembly, concept. We are looking for, applications. Using this assembly, concept. That are both good for data centers. Also, requires. Many optical pools. Hence a wavelength, selective, and space photonic, switch. Say wss. Might be a good option. Next, we will see the design, of the box. It's good to mention that we are using, standardized. Layouts, of indian falsified, and triflex, terms. Throughout the project. Here are the mask layouts, of two chips. The main optical function, will be provided, by indium phosphide. And the traplex, is utilized. As a carrier. On which the indium phosphide, will be full object bonded. All the two chips have specific. Dimensions. And locations. Of electrical, paths and of course optical, interfaces. Say the fingers, and funnels. All these, specifications. Can be transferred. To functional, designs. As i said for example, wss. Aiming to show the proof of concept. Now it's time to see the wss. It's a 4x4. Wss. Cross connect. This is the schematic. Field. It has four inputs. That found out, to a shuffle network. Followed by 16, gates. For broadband, input selection. And, four. Wavelength, multiplexer. With another, 16, gate, for wavelength, selection. Then a final, fan in, connects, to four outputs. This wss. Is based on the previous, work, done at due. However. It will be the first, time to try in our proposed. Assembly, concept. So we redesigned. The, wss. Using the standard, layout, mentioned, previously. As you can see, there are 16, four optical, connections. On each side of the chip in a four millimeter, range. When using 50 micron, pitch. It could be doubled, if we use 25. Micron patch. As for the wss. It has, eight inputs, and output ports. As you can see, light is injected. Into four parallel. Waveguides. And then found out to a shuffle network. Based on crossings. And cascaded. One by two splitters. 16, semiconductor. Optical amplifiers. Csoa. Are here, as case, to perform, broadband, input, selection. As for the azoid, base the case, they are electrically. Controlled. In the on state, the signals, are amplified. And in the off state, they are strongly, attenuated. Then the amplified. Signals, are transmitted. To the wavelength. Selection, stages. Each stage, consists, of a 4x4. Cyclic. Array waveguide, gradient, shooter. Say awg. And four iso, against. At the output, ports of awg. To, perform. Wavelength, selection. Finally. Cascaded. Two by one combiners. Are here. The fun, in the light. Connects, to the output, ports. It may be good to say more about the soa-based. Photonic, switches. The soas, can not only provide, faster switching. In nanosecond. Precession. But also optical, amplifications. To compensate, for unchecked, losses. Besides. I would also like to give you a favor, of other two designs. The first, one is a 40. Channel modulator. Array. Multiplex. By an awg. Onto, one single, output. This is also what, the schematic. Will show. Using. 40, of two places. The reason why we use after, places. Is the electrical. Absorption, modulator. Is a bad gap offset. For detuning. So we can't, use, integrated, lasers. Also the two photo detectors. And other outputs.
Are There. Because of, characterization. And reductancy. The second one is a transmitter, chip, with two widely, tunable, lasers. Microsoft. Is interested, in this widely, tunable, laser, for project, series. The chip is designed, to try the idea of nanoscale. Wavelength, switching. Tune the right wavelengths. While the other laser is transmitting. As i said, i can't go into detail, about all the designs. Since the time is limited. If you are interested. Please raise questions, after the presentation. The three designs. Of, indian, phosphide, pegs. Have already been submitted, to a multi-project. Wafer, run. On a material, functional, platform. Of smart photonics. Also. The triplex, tubes, and fingers. Are being fabricated. By lyonix. International. After the fabrication. Process, is done. The toolchip, will be, combined, by, an assembly, tool from falcon tag. We are collaborating. With these partners. In a joint. European, project. Eurostar. Project. We are targeting, to try the philadelphia, assembly, concept. In different, levels. In a dead level. To a bar level, and finally, a wafer, level. Then, a system, demonstration. Will be put forward. In the life of microsoft. Razer. Cambridge. All right. Let me close my talk with a brief, summary. Well, what i have tried to say today. Is a new assembly, method, for hybrid, pegs. By leveraging. Indian force flight technology. And silicon, nitride, based, traplex, technology. Especially. Using, the finger, based, flux, bonding assembly. For the hybrid. Picks. We are exploiting. This method, for emerging, data center applications. That needs. High density, fiber to chip, coupling. As i said, our objective, is to, demonstrate. A cost, effective. Assembly, message. That could enable. Many connections. In a relaxed. Alignment, tolerance. Three demonstrators. Ongoing. Range from light generation. To light manipulation. Especially. Optical, switching. I'm excited, to see, what will happen soon. Finger, crossed. So this is all i have for. Today. Thanks for your attention. Thank you very much. Wenjin. Uh, i believe, she is now online. Yeah. Yes. Excellent, so i can already, see a couple of questions. In the. Chat. Um. So let's start. From. The question from ben thompson. Uh he's asking, if. Uh the flip, chip, uh bonding. Of the, indium, phosphite. Die, to the triplet, carrier, means that. You no longer. Have to bring, all the electrical. Connection, to the edge. Of the, induced, indium, phosphide. Shape. And then. Can you instead. Do this electrical. Routing. On the triplex. Chip. Yeah yeah okay thanks for the question, uh as you can see savior. Has already, replied, it. Uh it says in our demonstrations. We have the predefined. Path the electric, pipe. On the, on the each side, the north and west side. North and south side of the turf. So we do need to route the metal tracks, to the predefined. Path on indian phosphide. And then through the philippine, bonding. The. Electrical, piston, in the phosphide, will be, as they bonded, onto, the trapeze, tube, and then, we will drive the indian foresight, chirp.
By The metal. Metal connections. Through the traplex. Yeah, i, hope this, answers your question, yeah, thanks for that i mean the reason the reason i ask is sort of having designed some of these trips before is you, and, you sort of spend, an awful lot of your. Precious, optical, chip area. Used up for routing electrical, connections, this seems like a really nice way of um. Getting around that problem and sort of heading towards, you know electronics, we've had multi-layer. Chips for, for many years whereas in optics we're still sort of an optical layer in a single electrical layer and i think this looks like a really, really nice way of getting around some of those challenges so yeah no thanks thanks really nice presentation. Okay thank you. Excellent, so we have an another. Question, from. Lee. Uh he. Is asked, the, uh, what what is the coupling, co. Efficiency. Uh coupling, efficiency. And what is the signal, stability. Since i saw it looks like, suspended. And can it be used, for, soi. Okay, all right thanks for your questions. And regards, for the first question. Uh yeah, as you can, as you can see i just demoed, the, mask layout, of three designs. So we, couldn't have the fabricated. Chair. In house, so we haven't do the. Coupling efficiency, now so i can't give you an. Exact number for this question. I'm sorry for this. And for the, second, one what was it the signal, stability. I. Don't know if i. Understand. Your question, correctly, maybe you can explain. More about this, i'm sorry for that. I don't know why you asked this question. Uh yeah sorry, uh, probably, i. I didn't, understand. Where, as well so i'm wondering. Because, this is transfer. Optical, signal from the fiber to the chip if i'm cracked. Um. Sorry transfer, what, amplifier. Or. Not uh uh, optical. Optical signal, like. Um. I think the main question, is, sorry. I can step in uh if, if that helps. Um. I think these these fingers they are they're like spring-loaded. So, um, they will stay in place because of the, of the of the mechanical, force that is uh that is holding them there. And then. They will not move even if the if there's a. Temperature, changes or, or a temperature mismatch, in the in the expansion, of the different chips. So the spring loading, keeps them in, in place so that makes the signal stable. The answering. Question. Yeah yep thank you. Okay thank you, yeah. Okay so regards, for the third question. I think in principle, the approach, can be, extended, to silicon photonics, tips. It's only required, that we. Adapt our design, of the. Finger and the page and of course the, funnels, on. On silicon. Silicon, chip, rather than the. False flagship. I think in principle, we can, but we haven't, tried this now, yeah. Perfect, and we also have. The last question, from. Nikolaus. He's asking. If. You have some values, either from stimulation. Or, experiments. For insertion, loss. So coupling, loss. And alignment. Tolerance. For the. For your scheme. Okay, thanks for the question. And. I think we, do have some. Experiments. Results, for the alignment, tolerance. Because we utilize, the finger connection. Uh not for this fluctuating, bonding but for, a kind of on-chip. Unchecked, probes. To connect, with the indium phosphide waveguides. And based on the previous, result. The alignment, tolerance, can be, as, plus minus, two microns. Which, which gives a, um. A relaxed, tolerance, for the fingertip. It's about 15, nano. Nanometers. I think this is what, we have, in previous, uh experiment. Is experimental. Results. But we have a still we haven't, done in the philippine, bonding. Concept, now. So i think we will give, more, uh results, and data, in, maybe fill the for the. Conference. I hope, yeah, thank you for the question. Excellent. I don't see, any other. Yeah okay, so, i don't oh i see a raised, hand. Is a, fotini, please, yes, just a quick question, uh i was wondering what is the maximum, number of fingers, and what is actually the limitation, for this technology, to increase the number of fingers if you could comment a bit on that. Okay. Thanks for your question. Uh, as you can see now we are using, the, 15, micron page, for the fingers. And as i show in the design. There are totally, uh. 16, four fingers. On each side of the chip, you can easily. Decrease, the page to 25, micron. Which promises, 40 connections. Per millimeter. That means we we can have 128. Finger, on each side. And, you can even, reduce the pitch, a little bit more. I mean to. For example 20, micron, push.
It Can give you more connections. But i i'm afraid like we couldn't do more because. The finger weight is, 10 micron. We still need need some space, in between, two fingers. So i think the maximum, stability. Is around. 300. To 400. For. Uh in a culture, in two sites. Yeah i think thank you yeah, okay thanks a lot very good talk thank you thank you. Thank you very much. Wenjing. I. Don't see. Any, other questions. On the. Chat. So i would like to thank you once, more, and move on to the third, speaker. So the, next speaker, is. Jing tao hong. Jintao, is a member, of the photonics, group at the university. Of, cambridge. Under the supervision. Of professor. Dapping. Dapping, too. Andreas. Georges, is. His, microsoft. Supervisor. His work. Relates, to pix. Elated, the spatial, light modulator. For phase and polarization. Modulation, of. Light. And. Today his talk is about. Phase and polarization. Spatial like modulator. Systems, for optical, data storage, in glass. Um. Jintao. Will present. Live. And, obviously, we'll be. Here for the follow-up. Question and answers. And, um, if, jintao, you wouldn't mind to, so scarlett. Will. Progress, your pre your slide. So whenever, needed, if you say, next, slide please. Scarlett, will be able to, progress, on, for your. Presentation. Okay thank you and the floor is, yours, jintao, thank you. So. Can everyone see my. Slides. Yes we can. Okay. So uh good afternoon, everyone. Uh my talk, uh topic, is phase and polarization, model, special ed modulator, for, fast data writing, on, optical data storage. First i would like to introduce. The. Application. The purpose of this, project. Writing data on glass. By 2020. More than. 39. 000. Eb. Data will be generated, by human. So, here comes the question. That, closer, amount of, data. How can we, store them. So, uh, there's, normally, five, criteria, for to evaluate, a data storage technology. Which are, first. Storage density. Second. Stability. And. Robotness. Third. Power, consumption. Uh, four. Red card, and excise speed. Fifth, the last one of course is cost. So. After. Thinking about this four uh five uh criteria, for, data storage, technology. If we store the. Data. On the using glasses. We can have the following. Three. Amazing, advantage. First it's a, ultra, high. Storage density. Because the data, optical storage using glass can provide you five, dimensions, of. Data, storage density. And also. Uh. Because we're using glass, so they will, give us a. Very low cost. In. Making the storage, medium. And also. Because of the high thermal and chemical stability, of glass, which means, we can use, this, type of. Data storage technology, to store. Code data. So. That's why there's a, a lot of brilliant, scientists. Start a project called project celica. To, realize. The, optical, storage, system using glass. In the beginning, i want to briefly. Introduce. The, project silica, writing. And reading data. Using glass. So. Writing, data. Uh we use a femton, second laser. Which can, create, multiple, pulses. In a very small area. Of the, in the glasses. This can create a nano structure. Which has a. Orientation. Applica, and amplitude. That two parameters. Is, how, uh later we defined, as the, beats, where, the information, was stored. And this part on the writing part is my work. Is focused on. On the reading side on the reading, data reading side, the data is reading use a polarizing. Microscope. Which can give you the. The. Birefringence. And the retardance, information. On the small, nanostructure. And this is the. Project silica. Next. Because this work is focused, on