ADI: The Collapse of the LIDAR Signal Chain (Full Presentation)

[Music] I'm now going to be talking about the collapse of the lidar signal chain and when I say collapse I am not necessarily predicting doom for lidar I actually mean collapse as in the integration of functions into fewer and fewer components throughout the lidar signal chain as from the introduction the company I work for is called analog devices we're a component supplier and we view ourselves as the bridge if you will between the physical world and the digital world and what does that really mean well we make a bunch of electronic components which perform functions such as sensing digitization of those sensing as well as of course signal conditioning and then connecting that information to the digital world whether that be the cloud or some other storage medium and we also support power management functions to sort of underlie all of these other more signal oriented or signal conditioning oriented functions and I think from this point of view we can offer a bit of a unique perspective to a conference like this because we don't have a particular dog in the race when it comes to which light our architecture do we think will or even should succeed our goal is to really support all of them and understand them of course to make sure we're offering the right support but through our components be able to work with all of the different lidar manufacturers just a little bit more detail on the group that I'm part of there's a business unit that we have called autonomous transportation and safety vide are is part of it but it's not the only sensing technology we also have groups working on radar and inertial MEMS just a quick background on those just to give you a little bit of context for where we come from in inertial MEMS we have a long history a little over 25 years ago we've released the first monolithic MEMS device for airbag deployment so that means we have a long history of developing products for safety critical applications and we'll talk a little bit about that later in terms of automotive radar we also have more than a decade of experience with products both at twenty four gigahertz and 77 gigahertz and today for automotive modules that are released and developed in cars more than half of them have adi content lastly for lidar though it's not as commercially available and deployed in cars as say radar would be for every lidar system at least that i'm aware we have some form of content whether that be data conversion amplification or power management okay so to talk a little bit more about lidar now I actually find it interesting to look back towards history a little bit and see what that might teach us historically lidar as with just about every technology has been used for smaller volume applications higher costs associated with that we can think of you know as originally developed for military applications just as with radar atmospheric monitoring wind lidar is common topographic mapping etc but moving forward we see integration reduction in size reduction in cost again as with just about any other technology leads the deployment into a broader set of applications of course with lower prices automotive that's the reason most of us are here but also robotics logistics smart factories etc are all utilizing lidar this follows a very similar deployment to radar where you may have started with large phased arrays here again perhaps initiated by the military but today we're sitting with small integrated low-cost automotive radar modules that are widely deployed if you go out and buy a new car today there's a very high chance that it has one of these modules are actually multiple of them in it and I think the interesting question is will we truly see the same thing for lighter we're all here to talk about this but are we going to get to these small sensors that cost $100 and are in every new car that's bought I don't really have the answer but I do want to talk about some of the trends that I think will help lead us that direction there was not really part of this talk I think it also brings up an interesting topic of POLST and eventually moving on to pulse compression and frequency modulated continuous wave that happened in lidar sorry and radar some years ago they're starting to they see the same thing in radar can we look at excuse me lidar can we look at radar and sort of extract extract or extrapolate from the progress and the path it took to predict what lidar will do I and don't have an answer there but it's sort of an interesting thought experiment okay to focus a little bit more now on the signal chain this is my view of it it's a little biased by being an electronic supplier I've probably minimized some of the photonics or optics components here but we have the transmit sync signal chain up top with a laser a lot of electronics to drive it perhaps something to steer it and then the receive signal chain down at the bottom where we have a photo detector array signal conditioning electronics and then some form of digitizer as well as processing so in this talk I'm going to focus on three areas that we see integration already happening or coming in the near future one on the transmit side and two on the receive side and those in particular on the transmit side of the laser plus the laser diode the focus there will be increased performance I'll talk a little bit more about that on the receive side we'll talk about integrating the photo detector with the amplification that connects to it and why you might want to do that both performance but also increasing the throughput through parallel operation and then finally integration of the data conversion with processing and here it's less performance focused and more focused on cost reduction and also reducing the heat which is a big issue particularly as you try to scale the size of light our systems ok so let's start by talking about the laser diode plus it's driver and before we get into the specifics of the laser I want to talk a little bit about the trend towards narrower pulses this I think multiple factors that motivate this trends to towards narrower and narrower pulses but the easiest one in my opinion to describe as I safety so that's the one I've chosen to include here this was covered a little bit yesterday there is this international standard six zero eight to five it actually constrains two things from an eye safety point of view at least for pulse slider though many people tend not to talk about both it constrains both the energy contained in a single pulse as well as the time averaged power that's transmitted which is effectively the product of the energy and a pulse times the rate at which you're firing the laser now it turns out at least in most systems that I've being able to look at pulse energy is the critical parameter again it's the product or sorry the power is the product of the pulse energy and the repetition rate but in most cases that I've seen the thermal load the heat generated by pulsing the laser very rapidly tends to dominate or limit the system performance before you actually reach the eye safety limit so in that case this average power limit tends not to be such a big issue and we need to focus mostly on the energy per pulse just to go through some numbers here the allowable energy per pulse is about 200 nano joules at 9:05 nanometers for reference that's about 40 watts of peak power for an ideal rectangular pulse now you can engineer around this to some extent as you use larger laser sources you can increase this by about a factor of 3 and then as you start to distribute the sources over areas much larger than the pupil of the eye you can get to even higher energies but just to keep in mind this is a reasonable frame of reference for where we're limited now when you're limited to say something like this narrower pulses are a good thing there are wind because for example if you want to maintain the 200 nano joules and you can shrink the pulse width you're not going to increase your optical peak power so you can see further now you might want not excuse me you might not want to do that maybe you can already see far enough in that case for the same thermal load for the same heat generated you can now increase your laser repetition rate which means you probably are getting better spatial resolution or perhaps even better distance through averaging and in both cases you're likely improving your range resolution with a narrower pulse there's necessarily less ambiguity in terms of time which translates to distance okay so we want narrower pulses and we want them to be higher what does that mean for generating this light well unfortunately well we want both narrow and large peak power that's not the way laser diodes and laser drivers circuits tend to work most designs limit the ratio of the peak power or the current to the pulse width so you end up with curves like this where on the x axis you have pulse width on the y axis you have optical peak power and for any given laser driver you can sort of operate on this line that's not what we wanted we said we wanted narrow pulses and a high peak power we want to be up here so why is that it turns out the culprit is inductance for those of you who may have taken electrical courses some number of years ago in college maybe you remember the characteristic equation for an inductor can be written this way it tells you that the rate of change of current with respect to time is proportional to the ratio between the voltage across it and the inductance current through a laser diode is proportional to optical peak power so what this really says is to get a big optical peak power in a short time we need this ratio to be big when we need to minimize inductance now inductance is not something we typically put in the circuits is just there as a parasitic it's part of the packaging that we use it's part of the board layout and it's even intrinsic in the devices themselves say the lasers or some of the other passes so optimizing the packaging optimizing the board design is key to getting the maximum performance that we want out of these laser drivers as an example if we were to start here and have the inductance we could either double the peak power we could half the pulse width we could do sort of anything in between or we could keep the performance the same in terms of peak power and pulse width and significantly reduce the loss because loss in these many of the lost terms in these circuits a proportional to voltage or perhaps even voltage squared so with operating with a lower voltage actually significantly increases the efficiency even for the same laser and laser driver just a little bit more details on why this exists we have to think about the principle of operation of a laser diode driver circuit the laser diodes here and the way this works is you store some energy in a capacitor and when this switch turns on the energy will discharge from the capacitor through the laser diode generating light the problem is that's not the way the circuit really looks and there are these parasitics and this inductance again it's lumped here together as one element but it's sort of distributed between this device this device and the routing between all of them is the problem so you can see that here your choice of laser your choice of FET the device down here the transistor as well as the capacitor affect the inductance the stack up of the board how you do your layout affects it and even using a raise of lasers many systems use force for lasers or sixteen or perhaps even more as you go to higher and higher channel count the practical routing limitations of getting between the individual laser say cathodes and the switches or the capacitors the practical aspects of that also impact the inductance and so we need to be very careful with that to back that up we did some measurements here so this is some actual experimental data the setup here as we chose an off-the-shelf a laser diode likely supplied by somebody in this room we use discrete electronics sort of off the shelf and happen to use a 30 volt supply and we measured this under two conditions we basically first did our initial design this is the actual measured pulsar time on the x-axis here in optical peak power on the y-axis this is our first design the blue curve and then we went back and looked at the layout we used the same components so we didn't switch the laser we didn't increase the voltage we didn't cheat in any way like that we were just careful about what is the path that the discharge current takes as well as selecting some of the passives such as the capacitor to minimize inductance and by doing that without changing anything else we increased the performance by 35% okay what does this actually mean well this means that being very careful but also having sort of deep knowledge about how these things are connected how you do the layout is is perhaps one of the most critical parts of getting the most performance out of your laser and that's something that I don't think a large number of companies are equipped to do so the next step in my opinion for the industry is smart partitioning of the packaging we're going to need cooperation for example between driver companies and laser companies to come up with integrated solutions that don't place this burden trying to get the green performance instead of the blue or perhaps something even worse on companies whose expertise doesn't really lie in that area of course it's not as simple as just minimizing inductance it also includes considerate considerations for thermal dissipation lasers get very hot you have to conduct the heat out of the package so the more things you put in there aren't necessarily better from that point of view and the impacts on the optics is well thermal lensing and that he might do as you increase it okay so now to move on and talk about the photo detector I'm gonna talk about the amplifier bit I'll use the term TI a for any of you who aren't familiar ti a stands for trans impedance amplifier most voted to have detectors I've drawn an APD here our current output devices so a trans impedance is just something that takes a current and turns it into a voltage and that's typically done with a resistance placed across the terminals of an amplifier in order to efficiently draw the current out of the photo detector here and turn it into a voltage at its output so what are the interesting or important parameters of at EIA for the for its application in lidar the first one is bandwidth so we were just talking about narrow pulses and trying to generate those well it turns out in the receive signal chain you need to match the bandwidth of your amplifier actually the entire signal chain to the pulse width in order to get those benefits and you can run through some math but you find that roughly the bandwidth of your receiver should be greater than one over the pulse width so as an example for the five nanosecond pulses we were talking about you need about 200 megahertz of bandwidth how do you figure out what the bandwidth of at EIA is well the easiest way is to go to an application note and find an equation like this it's proportional to a few terms something sort of intrinsic to the amplifier itself called game bandwidth product the resistance and feedback here and then the total capacitance at the input which is a combination of a bunch of things but does include the capacitance of this photodiode state-of-the-art trends impedance amplifiers have something on the order of a gigahertz of game bandwidth product so that means to achieve this 200 megahertz of bandwidth we were talking about you can use something like a four kiloohm resistor if you have one Pico farad of total capacitance at the input now if you want to go to narrower pulses and you want higher bandwidth one of these two has to change you for example can use a lower resistor but that's not good because it increases noise and so it's going to be critical to minimize this input capacitance in terms of achieving the maximum bandwidth to talk a little bit more about noise I grabbed this diagram from a datasheet it shows the noise spectral density so the current noise at the input and the y-axis as a function of frequency on the x-axis and it shows it for several different noise contributors the red line here is the combination of them all but you can see that the feedback resistor the one we typically think of is this I guess yellowish or orange one here it's flat with frequency but there's other noise terms you have to consider and especially consider when you get to wide frequency or wide bandwidth systems because they're noise density increases with increasing frequency and one of the particular ones of interest or of importance if you will is this thing that's labeled en that's the voltage noise input referred of the amplifier and the reason that's relevant is to turn a voltage noise into a current remember the APD's or the the the photo diodes let's silicon photomultipliers whatever you use their current output devices so we want to refer all of our noises to a current you do that by multiplying by the input capacitance so if you have a larger input capacitance this entire green curve or blue curve I guess it's blue shifts up vertically and that makes the noise worse and it's actually a double whammy because not only does en increase with input capacitance but it also depends on frequency so as we want to go to white or bandwidth Ian's doubly painful you're integrating more of this noise that's out here and it becomes even more dependent on your input capacitance so again this is motivation for why minimizing the input capacitance is critical this is part of photodiode design right we want the photo diodes themselves to have very deep junctions that are fully depleted to have as small capacitance as possible but it's also a function of packaging and layout just as it was with the laser drivers and this is going to lead us to integration now something I haven't talked about too much is the arraying up of detectors but I think that is clearly a transits as we talked about for channel is 16 channel lasers it's very common to see 16 channel 32 or even higher channel detector arrays eventually limit I think well for one export considerations but in either case we're moving to detector arrays and that means we're moving 2t i-a arrays as well of course as I mentioned the detectors need to be optimized for low capacity but beyond that we have to think about the packaging so what I've shown here shown here is a a package the right side is sort of a cutout of it where you can see a photo detector array on top of a TI a array this happens to be a backside illuminated voted detector so it's flipped over illuminated through the back and it's directly bumped to the TIAA array that that minimizes capacitance that minimizes the parasitic between those two connections or those two components rather and optimizes performance now if you do this you have to think about low-power design because you don't want to heat up photo detectors or the dark current increases so low-power design becomes even more critical the last thing I want to point out is the people in the world who make the best T IAS are not the same as the people in the world who make the best photo detectors and so getting the best performance out of a package like this just as it was with lasers is going to rely on collaboration or partnerships around our industry and again this is another instance but I'll talk about it more a little bit later so the last topic to cover is a 2d conversion as well as its integration with processing I want to look at the return waveform so as a function of time what signal is coming into my lidar receiver and I'm just sort of drawn it as this noisy waveform here and now we have to decide if something was there this direct time of flight was covered a little bit yesterday and the phantom intelligence talk but just as a recap in case you weren't there effectively what you do is you set some threshold and you wait until the signal crosses over that threshold and when that happens at that event you decide or measure what the time and therefore what the distance of the object was of that crossing now as you can see here well it may be a simple power and cost efficient solution because TD sees themselves are relatively small and I should say not not particularly complex as well as the sort of detection scheme itself being not particularly complex the SNR requirement is going to be higher you need this threshold to be relatively high so that any noise that's down here doesn't actually accidentally rather rather trigger against your threshold so that means systems built with this direct time-of-flight approach are in my opinion necessary to be limited to shorter range than they could otherwise be so what's the alternative we can go to full waveform digitization we can take what we had up here digitize it with an ADC we generate many more samples but now that we have all of these samples we have the ability to do certain things such as filtering and that reduces noise so I've drawn a much smoother waveform here this would be after filtering as compared to here or there and that maximizes your SNR particularly in the case that you know the waveform you transmitted you can do a matched filter which can be shown to theoretically optimize SNR and so these systems are going to be longer range systems if range is your problem I think you're gonna do better with full waveform digitization than you are with direct time-of-flight but it doesn't come without penalty there's overhead associated with generating all of this data both in terms of the power but also moving this data around there's a lot more data over here there's one data point here when did it cross here there's I don't know what I've drawn maybe a hundred samples something like that in amplitude values associated with each of them going back to a signal chain this is a little bit redrawn from what I had earlier but a similar thing where you have t.i is connected to a dcs and now i've explicitly called out the high-speed serial link between the ADC and the back-end process or in this case an FPGA which i think is at least for prototyping purposes the most common commonly used processor and let's look a little bit more detail as to those penalties so with the high-speed ADC they're not cheap for something that's running at a Giga sample per second or so you're looking at probably a hundred dollars per channel in what I would consider reasonably low volumes 1,000 units they also consume power half a watt so if you have a large array of these you're you're spending hundreds of dollars as well as several watts of power what about this length is high-speed serial link one example would be jst 204 B that's a standard Suri's high-speed serial link and the thing I want to call out here are there's actually two things one is there's a lot of power associated with it so again you're spending a watt just to move information from here to there you're not doing anything with that you're just sending it from point A to point B and incurring a penalty but perhaps more significantly it limits the choices of FPGA FPGAs that can accept this high-speed serial link are not as cheap as perhaps ones that use lower lower bandwidth links particularly when you need automotive qualified FPGAs so if you actually look at what those FPGAs are you probably are looking at more than $1000 so there's a lot of cost here several hundred dollars here several thousand dollars there from just two components in your lidar system you haven't even really built the rest of the lidar system but you've already completely blown your budget at least for the low low cost 8's stuff over and over again and it's effectively limited by this high-speed serial link so what can we do well let's go back to the ADC data waveform that we had before again it's high speed but the thing to recognize here is that most of this is not an echo if I were to take a rangefinder and shoot the back of the room I'm probably only gonna hit one thing that wall back there everything else is just space there's no real way data of interest coming back to me so wouldn't it be great if I could just recognize these two areas that I previously knew because I happen to draw this waveform that were returns if I could identify those and send only that information it would really help me out and that's the concept here let's identify data of interest through some mechanism let's identify where these two returns are and transmit only this data and ignore the rest of it what does that actually mean from the throughput point of view it's actually though it's a simple concept very effective so I ran through an example here where just to keep the numbers around I said what if my lighter needs to see 300 meters that's two microseconds of the digitized data again using the example of five nanosecond pulses that I've been using I'll allow for eight possible echoes I think eight is probably more than you need but let's allow for some false alarms that means you only need to transmit 40 nanoseconds out of the full two microseconds of data that's a 50x data reduction okay 98% of the data that my ADC converted I don't actually want to see so if I can get rid of that that really could have a significant benefit since I've talked a little bit a few times about automotive qualification I want to bring that up here it's come up a few times and different talks and mostly in the context of temperature and it is true that temperature is important whether that be 85 see for example for your lidar typically translates to 125 C or even higher for the components that analog devices or like companies would supply that's only a part of the story and it's actually a small part of the story even though it's the part of the story I hear people talk about the most automotive qualification really is a lot more than temperature range I've listed some things here apqp is a section essentially a set of processes that we follow to increase the likelihood that we can meet the automotive standards but I've listed some more of the AEC automotive electronics council qualification requirements up here that are beyond temperature so we have to think about things like all of the other stresses high humidity as well as high temperature making these things reliable and last for 10 years in the field 10,000 hours of lifetime and to do so with a very low failure rate the components we sell that are Automotive qualified have a target failure rate of less than one part per million and I would I would ask the lidar developers here if they truly believe that the systems they're building today they could build a million of them put them in the field for 10 years and get less than one return on average it's really quite a challenge and that's before we even start to talk about functional safety functional safety is built on top of all of this you have to have done a good job follow the automotive development practices and then you layer on top of that safety elements I won't really get into this that could be its own talk but I think it is a significant challenge that's going to face all of us and currently being significantly underestimated in terms of how much pain is going to cause a lot of the the ecosystem in the coming years in just my opinion okay speaking of the ecosystem this is the last thing I'd like to talk about I'll first talk about how I see all of the the sort of different players interacting today and then how I think that's going to change in the coming years I'll first state that while I put some logos up here this isn't by no means an exhaustive list please don't be offended if your name's not up here I just pick the first year that popped to mind or we easy to find the logos of so we have users up here effectively tier ones and OEMs and disruptors and then sensor developers and today I think generally the specifications flow from users down to developers who respond to those of course and then below that we have where my company sits broad market component vendors we have relatively standard or general purpose components that we supply into the sensor developers and then I've separated out here specialized component vendors who have in many cases unique technology often specific to a particular type of system whether that be FM CW or some particular aspect of scanning or processing or whatever but in this case they're they're application-specific so I think this is reasonably well established and we're all pretty comfortable with this in the future what's going to change so the first thing I've done is I've separated out the vehicle developers from what I'm calling the industrial design of the lidar and I've done this because I'm as you'll see in a second I'm not going to have an arrow between these two and why is that well we talked about automotive qualification I think we're already starting to see it but I think it's going to be very commonplace or become commonplace that the vehicle developers up here are going to demand basically accept accept no substitute for the experience these guys have in industrializing designs for automotive all right so we're not going to have direct in Iraq I mean sensor developers will continue of course to talk to vehicle developers but we won't have sourcing of sensors directly through this and in that case the sense of developers effectively become technology development arms of tier ones and you're already starting to see this happen with a lot of partnerships and announced I think that will continue even some acquisitions I then merged all the component developers into one big box at the bottom and the reason I did this is I think the components that are sourced to the sensor developers will become much more application specific we won't have as many general-purpose components flowing into these applications we'll have designs developed uniquely for them of course more partnerships I brought up several examples of that sort of within this community you know company a talking to be talking to C to come up with these these innovative solutions I also think many of those solutions will flow up directly to vehicle developers we're already seeing that where these guys reach down to the bottom to find technology that they find compelling and then direct buys or direct design ins to the tier ones and the sensor developers and the last thing is you'll notice I don't have the small component technology companies listed here anymore I didn't want to call anybody out in particular but I do believe that in large part they're going to go away that will in some cases be extinction if you will I know the cases it will be absorption through mergers and acquisitions but in large part I think this base is going to consolidate as well as some other consolidation elsewhere ok so to wrap up we talked about performance being needed we went over integration of three areas that I think are going to integrate in order to drive performance we talked about the industry evolving to no longer have just this sort of linear relationship from top to bottom but a much more interconnected web between the vehicle developers the sense of developers in the component suppliers and of course no single company can do it alone so I'm looking forward to our company is looking forward to being part of this and hopefully with the help of all of you [Music]

Channel: Analog Devices, Inc.

Views: 2,106

Keywords: analog devices, adi, adc, dac, converters, blackfin, mems, amps, amplifiers, power, isopower, icoupler, digital isolators, microcontrollers, embedded processors, analog to digital converters, digital to analog converters, LIDAR, LIDAR System, LIDAR Signal Chain, Automotive LIDAR

Id: z2IOrgMyUz4

Length: 30min 38sec (1838 seconds)

Published: Thu Nov 21 2019