a kind introduction so thank you for kind introduction and today i will uh give uh another version of my doston slides it's improved with with several new achievements which we have done in uh recent years so uh firstly i will i will just uh give the outline of the talk so it's it's quite long but i will try to skip very deep technical details so just mainly focus on what we are what uh what what is most important takeaway so i will first tell it about about myself then about motivation and some basic principles which we will need to understand uh so we can get through the technical material then i will tell about the spec specifically about the device we use to achieve these all results and then we'll show several experimental demonstrations which we did by using different signals and then i'll show what we plan to do in the future with this technology and some conclusions so let's start first with education some and external external states so i did my last degree was docent in physics with specialization optical communication but all three degrees i have done in riga technical university i have been also partly in uh technical university of denmark and then i came to sweden back in 2015 so and i have been traveling quite a lot and these are external states which means that i spend some time for example most of these visits when i was employed at rice then i spend some some days at foraying the companies and most of these visits are related to the component i'm going to talk about today so and then yeah and then the motivation uh so i i give a example from myself so uh i was thinking how many connected devices do i have at home and then i was started connect counting how many they are like including printers tvs old mobiles from from children and i get close to 20 connected devices which was not like that some time ago so which means that uh you consume a lot of data and then this data has to be stored somewhere and transferred to you quite quickly so uh today i'm gonna focus on now i don't i'm not going to focus on all the network which exists and which parts are there but i will most specifically look at the uh for in data centers and these links in the data center inside and between data centers as well so just there's some examples of course there are different technologies available for uh trans okay transoceanic links or metro links or 5g front hole and access networks so but the main focus will be on on data center interconnects and you probably know what is data centers and they know that they consume a lot of energy but they store also a lot of data and then i have a nice example here which was shown in eco 2021 conference in in one of the opening speeches by researcher from nokia bell labs and then here here we see a picture uh and these are the view of data centers and you can see that they and then there's also a date when it was built so they built one in 2009 then 2012 14 15 16 and in 2017 they had to build two data centers so this can go exponentially so you don't want to build another 10 data centers build 10 data centers in the same place even though microsoft is already testing out how to uh put a data center on the bottom of the sea so uh and and and how to overcome this is to make the devices which are used for communication in the data center uh higher speed and higher energy efficiency so you can transmit more information with the same energy so uh and now the thing what i'm what uh what i'm talking about it's it's technology it goes into these packages it's a specific packaging for transmitter and receiver for a pair so and then it goes into some equipment and these are what we typically been dealing with in data centers but then there's also one uh trend where the when this all components goes into here and then there's only fiber connecting this place to here and it calls on board optics which requires a different mindset mindset how to create how to integrate uh different different parts of of of the transmitter and receiver so what uh yeah i will i will talk about some things which are inside these modules uh and which are used for for high-speed communication so uh basic principles so i'll just show all the pictures so um so what what do we do in in a physical layer so we want to transmit uh data so from uh from point a to point b so uh you have a lot of zeros and ones which you want to transmit and how do you do that first you represent logical information into digital signal and then uh this digital signal for in order to transmit it over some distance you need to modulate on some carrier and then you have a carrier in this case the carrier is uh it's a it's a light wave carrier which has a certain frequency and in in optical communication that's in terahertz uh the frequency and in wavelength it's uh this particular device is around 15 15 nanometers wavelength and then you see when you have uh ones and zeros then you have a you have the light or you don't have the light and this is a time representation of the signal so and then what happens over the transmission that you don't get exactly the same signal back so you get noise impairments and different other impairments coming both from the components and also from the media so this is a classical setup which which have been which i will show throughout the presentation so we have a laser which generates a lightweight carrier then the modulator where the modulation happens and then we have optical fiber over which we transmit the data and then we have a photo detector which down converts from optical carrier back to a baseband signal and then you get noisy but then with specific signal processing routines we can recover back the same data and we usually compare the data we sent with data which we received and then we count how many errors did we get and then we divide number of errors with the bits transmitted so we get something for very reliable transmission then you get a bit error rate we call it and it's down to ten to minus nine if you don't use some specific signal processing and uh what we do in our experience we achieve something like bitter rate of 3 e minus 5 or or sorry 1 e minus 3 something at that level so i just i would just want to introduce some basic principles and then what how do we represent that so we take all these ones and zeros we take only one period and we put put one on top of each other so we get something like eye diagram so when this is a zero level and this is one level and this is a transition from zero to one and you kind of get an eye opening in in experimental systems this eye opening is not that wide so there are also tricks and methods how to make this opening wider and how to distinguish whether you have received zero or one and of of course you can what we want to do is as i said we want to transmit more information with the same amount of energy so and then we try to implement on the same uh bitrate we put more levels and then we can say that this is one symbol this is second this is third and this is fourth symbol and each symbol corresponds to two bits so which means that we double the information we can transmit at the same uh baud rate so yeah and then practically it looks something like that so uh and then there's also some tricks what you can do because uh one one more one more uh thing we need to know throughout the seminar is that uh this is a time domain representation of the signal but we can also look in the frequency representation of the signal and that's where the bandwidth comes into the play and there's some tricks we do when we have a limited bandwidth in the system then we can do a trick called pulse shaping when we can reduce the number the the amount of bandwidth we need so this one uh so we this one is the same signal but we apply a pulse shaping here so we need less bandwidth for this signal compared to the signal so we lose in horizontal resolution but we gain in vertical resolution and that we can do also for multi-level signals like for example for dual binary and also for uh pum4 where you have a three opening so but there's also price to pay obviously if you have one opening then the distance from here to here is larger then you have the same distance divided by two or the same distance divided by three so there's a penalty to pay so and uh and obviously you need to optimize depending on what system you have for example here i have a time domain representation of the signal and here is a frequency domain representation of the signal and for example if we use a dual binary then we need a half of the spectrum uh to transmit the same amount of bits but then we lose in the resolution so uh then we need to use these two eyes instead of one big eye so there are there are things you gain and there are things you lose so and then we try to run a lot of numerical simulations comparing these non-return to zero for example when we have a limited bandwidth so this is 112 gigahertz and this is 56 gigahertz and we do it for non-return to zero and then we use a dual binary approach and on this axis we have received optical power at a certain literal rate so uh which means that if this one is is the lowest this is it's the best because then you need the least power to detect the information and we see that by decreasing the bandwidth in both cases for run on return to zero or for dual binary uh this this curve goes up so we need more power to the to to reach the same bit error rate to to get the same number of errors per different for for a fixed time period so uh and then we can we get the same results also for uh palm four when we look at the band with limited systems and yeah and also if we compare if we compare dual binary and pump four for dual binary we need even more so we need the most power for dual binary which is filtered and the least power we need for non return to zero which is not filtered so uh and then now we know the bit error rate we know the bandwidth we know the zeros and ones and time representation now let's see what is the practical device we can use to generate the signal in as this is the part of the laser and modulator so and these externally modulated lasers have been developed in kth since 1997 and then this particular is the outcome of hecto project which was led by professor ruben westergren and here are some references to original papers where where they describe how the chip works what is the principle so basically have a laser part here and then you have a modulator part here and it's a bit special track structure which is called traveling wave and using this structure you can extend the bandwidth of the device and in this case they claim they expand up to 100 gigahertz so on that this is the how the package device looks like and this chip is is inside there and this is the hundred gigabits they managed to achieve uh during hecto project back in 2010 and uh when i came in 2015 this device was laying on the shelf and uh i was thinking okay if you can do hundred gigabit of king with this one and you have 100 gigahertz of bandwidth so you basically don't need that much of bandwidth 400 gigabit as we i was showing numerical results before so thinking okay can we do something with it can we can we transmit higher speed than just 100 gigabit with this particular device and here here are some basic characteristics and it has quite high output power which is also very important if you don't use any amplification so you just have fiber which has lost you you need to compensate for that so then we started the first experiment say okay let's try to do something at 100 gigabit but let's look at dual binary approach and this experiment was done in uh ghent university together with the researchers from from there and also from some chinese university and then there's also one more concept we need to understand that there's also a uh an equalizer so what the equalizer does it's it's predicts the response the frequency response of the system and in this case it's a compression so you have the transmitter you have i i opening then after transmission so you have a bandwidth limitation in the system and it's also noisy you don't get any eye opening so if you apply the equalizer then you can get opened eyes of dual binary so in this case we had that we were severely limited by uh test boards for the transmitter and receiver and also rf amplifier so optical components were not the main bandwidth limiting factor here and that you can see from this picture so this is all system this is a frequency and this is the amplitude and then you it's combined the rf photodiode rf amplifiers and then it's this one is fiber photo diet and eml so you see that for one for 500 meters you get the bandwidth very flat up to 70 gigahertz and then when you start to transmit over fiber so the fiber introduce not only uh attenation but also some frequency selective fading so which becomes worse with the with the distance so basically this having this dip into the spectrum is not very good so i will show what we could do about it and here's also some pictures of uh how the devices looked like [Music] so and then we showed one case so in this case the system was bandwidth limited and then we tried to look at the speed of 70 gigabit per second because we wanted to measure the beta rate down to ten to minus nine and this is the like received power what was required so for example to get to uh ten to minus nine uh you need it like around uh minus 0.5 dbm of power and then when we increase the distance of non-return to zero then then you see that we cannot reach the the to very low bit error rate so it's very high and then if we transfer it and start to detect instead of the middle of this eye but we look at this eye and this eye instead then we can then we can gain the same bit error rate and we can gain like 5.5 db uh at this level so and we can get back to 10 to minus 10. so there's obviously a benefit there by using dual binary when when the system is badly limited and then we did the same uh measurement also but 400 gigabit and then we were measuring dual binary ice and then it was we achieved a higher bit error rate but then we managed to transmit over uh 100 gigabits over two kilometers and the problem here is that the the laser which is emitting at 15 15 nanometers wavelength and that's for standard single mod fiber the chromatic dispersion is pretty bad so you cannot go quite far so yeah but that can be changed in in the device design so then the next one was on off keying we tried uh and for that reason we went to uh we want to went to uh france and we visit at three five laps and they have a a specific chip which can generate on-off keying signal with a very good quality at very high speed so at that time when we did this experiment that was the highest rate of on off king even though it was quite close die but that at that time was the highest rate and we managed to uh publish this work at the ofc conference in post deadline session back in 2018 and then we used this signal at the transmitter and here is how the transmitter look like so here here uh here is the amplifier we used also development hecton here is our laser plus modulator and then here are some references to to people who who did the made the devices in in the hector project and then what we what what we did so we needed to generate this high speed signal and then we started with 51 gigabit per second and then we delay an interleave so we use a two to one selector as a as a package component so we doubled the bit right here and to double it one more time then we needed to do it on the chip so uh the package of the component actually is affecting how high speed signal you can send uh sense with it so we have to use a selector on the chip and then this is the signal quality we get at the at the output and yeah and the biggest problem here was actually the clock distribution to synchronize all these uh steps and you see that this there's a very nice thin sine wave and this is wider sine wave so this was actually setting the limits on on how good we can go and that was related to some 10 megahertz reference clock quality so okay and then we did this for inter data center distances we did it for 10 kilometers and for 80 kilometers and this is how the eye diagrams look before digital signal processing and there's quite heavy signal processing done some heavy signal processing routines uh employed and then we get the eye opening 140 and not very open eye diagram at 204 gigabyte per second but at that time uh that was uh the best one could achieve and then i'm saying uh because we have some newer results which i'm going to show later on and then we show a typical bit error rate versus received optical power and in this case we're talking a bit rate down to 10 to minus 3 and then we need to use some post processing like forward error correction which we which we which is not very acceptable in in data centers because you want to remove as much heavy signal processing as possible and then we did the same bit rate measurement for 200 gigabot and this was done for transmission over 10 kilometers of fiber and in both cases we were we were compensating for the fiber dispersion and these are compensated which are made here in in sweden in shista there's company called proximian which make uh fiber brag ratings which we used in this experiment and then what we were looking for was can we get uh multi-level signals into the into the device even more than dual binary so then we visited keysight and that's one of the trips and here you can see that we had the access to arbitrary waveform generator which can generate these electrical signals and we have a digital storage oscilloscope which you can store the signal on your computer and then you can do a processing on it using matlab and here is the component of the transmitter and the component at the receiver and then we had some amplification and and attenuators and and and stuff so at that time that was the highest rate arbitrary vacuum generator and digital storage oscilloscope which was provided by keysight and we use the same [Music] electro electro absorption modulator in in in this experiment so and then here here's a basic sim setup so we have a laser glass modulator here then we have an electrical signal generator we amplify it then we trans transmit it over 400 meters of fiber and then we compare two scenarios where we amplify signal or we don't amplify the signal and this was one of the first experiments where we managed to show that you don't need to amplify if you go if you go to multi level so four hundred gigawatt from four so and then we also look at even higher uh uh levels of amplitude for this one so uh and here you have a typical response uh frequen saw the bandwidth available in the system and the main limitation was coming from arbiter away from generation generator which was up to 55 gigahertz but then as you can see the green curve and the blue curve they follow each other quite nicely so these dips here were coming from arbiter away from generator and neither the less we managed to achieve a 100 gigabyte uh pum4 so in this case we have a three eye openings we have four levels and which means that the 100 gigabot stands for 200 gigabit per second so which means that we also achieved uh what we could do with dawn of king so we we doubled the rate of trans signal uh compared to uh compared to a previous experiment we double compared to hecto results we managed to double the bitrate we transfer transmit with the device and then we get the same i i also after the transmission and in this case we didn't use much of amplification or or post equalization in in the system and then we also compared the the bit error rate and we optimized that for two modulation formats for palm four where you have three eyes and for palm eight where you have seven eyes and then for palm eight the eyes are so tiny that you can achieve bitter rate of like four point e minus two which is very high bitter rate and not acceptable in in data center interconnects so we then we focus more on on pump four and then we also look com add amplifier and see how good we can be with the amplifier and then we see that without the amplifier we are not that bad in in the signal of performance and we did also that for 300 gigabit uh on from eight so we yeah so it was just more bits on the same symbol and then i would like to show the newest results of the of the which was recently uh shown that the ofc conference 2022 just some months ago and then so as i said you i said try to look at the eye diagram which we had before so now we have an open eye diagram for 200 gigabytes of king and that's without efa and then as as a state of the art so this is the experiment also showed bit earlier today so that was the first experiment we showed 204 gigabyte of king the same group but which three five laps they they were collaborating with the with polymer polymer uh polymer uh modulator group who uh from jurgen luthor so they they managed to achieve a higher speed 222 gigabit per second and the benefit of that modulator that you don't need to have an amplification stage here which yeah but then there are limitations on how high power you can put on it and then there's a quite large loss you have two erbium dot amplifiers in the setup to transmit over 120 meters so on the eye diagram is not so open then the next work was also shown in in recent eco conference so it's two premier conferences in optical communications one is ofc which is usually in us and eco which is usually in europe so uh and then they showed also uh 220 gigawatt but the eye diagram was closed so and they had to use edfa and they used also polymer modulators uh here so and then why one very good competitor to our externally modulated laser is same film maxender modulator uh in this case they didn't use any amplification or at all and then they they have to use an external uh external laser which is not monolithically integrated with thin film alexander modulator but that's that can be solved and that's i've seen recent papers where they integrate the laser together with the model later so hopefully we'll see some new world records coming soon so uh and then here is the setup with the picture it's pretty similar to the setups i showed before we generate the signal and in this case the arbitrary waveform generator was with improved bandwidth and improved sampling rate so we could generate higher speed signals with it and then we use the same external modulated laser the fiber photodiode and also the oscilloscope where we record the signal and and store it so in this case we managed to achieve quite high output power because we we apply different settings on the modulator and then we drive it at 17 degrees celsius then we reduced the bias to 1.6 volts so which
allows us to get higher output power and yeah so here you see also the arbitrary waveform generator on the picture and i have a component here which is the laser and photodiode which is here and then we have a full spool of fiber in between and optical attenuator here and these are just the control instruments to provide required currents and voltages to to the device and then we have a computer on which we can obviously on on digital storage telescope we can also do some processing and we compare this eye diagram with the eye diagram we can achieve by using our digital signal processing routines we have been creating uh here in sweden since since 2015. okay and then uh coming to the results so it's the typical bit error rate versus received optical power and then we do this at two transmission speeds one is at 170 gigabyte another is at 200 gigabytes so and then uh what is acceptable in intra data center communications is that you receive a bit rate which is the limit of hard decision for the error correction so in this case was 3.80 minus 3 and then and then with 107 gigabyte we could achieve better signal quality but of course lower speed and here i also show uh the eye diagrams of 170 gigabot like before transmission after transmission and 200 gigabyte before transmission and after transmission and then so in this case we showed that we can achieve 200 gigabit per second with very wide open eye and then we did optima and as i said we did optimization with some equalization in in the receiver and we the main finding was that we need some of some uh it was a special structure of the of the equalizer it's called uh decision feedback equalizer so we needed some feedback in the equalizer to get the signal quality uh better so and then we also did the uh the signal uh we we try to look at what is what happens if you use 100 giga instead of instead the 200 gigabit we get to same 200 gigabit but using pump 4 and 250 gigabit if we use palm 6 so we use the same setup the main difference was that we changed one amplification amplifier in the receiver so but then we managed to achieve quite good performance the the the speed in in this case for pump 4 is 200 gigabit as well but then you get smaller ice but the bit error rate is this is also crossing uh hard effect limit and then we need uh uh for that we need around 0 dbm which means that if we have 3.3 dbm then we
can then we have a positive power budget so we can transmit uh more signal uh over it and then uh yeah so and then we also did the same with palm six then you see that there's even more ice but in this case we were not really limited by bandwidth uh for for this baud rate so we managed to show that multi-level signals also work and then yeah so uh and then all this works i didn't do just myself there's a huge list of contributors uh on this optical interconnect topics and we are not stopping here we do also plan to do following up experiments to get to get even better results and what i wanted to also show so i have showed that this part of the of the world record which was 200 gigabit on off keying and we also showed the 11 gigabit on four with quantum cascade laser uh so uh directly modulated it's it's much lower bitrate in this case but it's a new wavelength which we use for free space optical communications and in that case that was also a world record and also accepted in yfc conference in post-deadline session which doesn't happen that often to have two papers accepted in the in the same conference in in a post-deadline session so uh yeah so these are the highlights for she from cheese to high-speed transformation lab and then i have a bit or to tell about twilight project so we come back to eml and then uh we have a package device which offers a single channel and so what do we want to do in the future we want to integrate those devices those several several devices to get to 1.6 terabit per second and that can be done for example if we do 200 gigabit per wavelength or or even higher then we need a number of devices integrated side by side and that we plan to do in a twilight project so and yeah so so hopefully you will see some arrays of externally modulated lasers running of 1.6 terabytes per second in in a soon future and then i want to read all the conclusions so they're quite technical but i will just would like thank you for your attention and then i leave some space for questions
2022-08-25