GaN PAs for Microwave Line of Sight Link Applications

well good afternoon I’d like to present you some work I did with my colleagues Stuart Glenn and Tony Richards gallium nitride technology sure you all heard gallium nitride it’s got proven capability for high power amplifiers now a lot of the early applications of gallium nitride they focused on high levels of saturated power in fact people seem to go out the way to develop as much saturated power as they could and it is a more expensive technology than than gas or in silicon and people would tend to bolt a game nitride booster on the output or of an existing solid state PA or sometimes they built a dome nitride amplifier on the import of a twt and it was all about saturated power however as the Machop as the technology is matured things are moving on and gallium nitride is increasingly adopted for highly lity linearity applications probably the most well-known is base station power amplifiers and there’s a lot of work underway now with people developing dilemma i tried transistors to replace the current LD Mo’s transistors and even freescale and nxp the biggest LD most players are both heavily investing in the gallium nitride market I’m going to talk to you about gallium nitride pas for line of sight link applications and this is a market which is currently dominated by gallium arsenide we’re going to start by looking at an example a a PA for 15 gigahertz line of site links and this this is a European bun there are various buns from round about six gigahertz up to 40 gigahertz we’re going to develop this PA using crees quarter micron garland silicon carbide process the initial design will optimize for the conventional measure of linearity third-order intermodulation point but that’s not what is really important for a line of sight link what’s important is that when you amplify a complex modulation spectrum that you preserve the modulation fidelity so we’re going to pass 1256 through that PA and look at the soul of degradation we see we’ll look at spectral regrowth the adjacent channel power ratios we’ll look at the constellation diagrams and the error vector magnitude which is a well used metric for constellation distortion and we’ll look at how these parameters vary with the bias of the amplifier so this was the initial target specification of our PA so the frequency range of 14 and a half to 15 point 35 and this is the the so-called 15 gigahertz point-to-point band again something greater than 20 dB it’s going to be a packaged IC and a gain of 20 to 25 DB at microwave frequencies is a sensible level for a packaged IC output ib3 we were aiming for 445 DBM at 22 DBM of output power now one of the reasons for licking at gallium nitride is to increase this IP 3 level and the reason we want to increase the IP 3 we’re not likely to generate more output power in actual fact the knee incumbent gun at gas solutions were looking into about better linearity so that we can up the level of the modulation the order of the modulation quam 256 is now routinely used Quan 512 and thousand and 24 is being rolled out and people are investigating higher order modulation schemes now as you’ll see as you go higher and higher all of the constellation points get closer and closer you can upset that constellation with a much lower linearity a much lower level of distortion so you need much higher levels of linearity in order to amplify it the saturated output power 38 d BM and pa e for the whole amplifier of thirty five percent and one other important point lot smaller area

than gallium arsenide parts now gun per millimeter squared is more expensive than gallium arsenide however the power density of gun is higher so you can get a smaller chip we’re using the creek or two micron down on silicon carbide process we’re going to buy us it or put 28 volts VDS and let the frequency range we’re operating it has about 4 watts per millimeter of RF output power this is one of the transistors and it has these inter source virus and these are nice for keeping down the source inductance which helps preserve the available gain from the transistor that I our microwave frequencies they also spread the gate fingers quite nicely which is important for thermal operation and that gives you this wide gate picture the PDK that Cree supply is nice it’s got a facility in it to predict the thermal effects of the on the RF performance you have to put in the thermal impedance of your transistor which is down to the designer to work out so it does require some skill but given that if you type in the base plate temperature then the model will tell you how the performance varies due to the self heating effects now using these models and our CD simulators we can’t tell you exactly what the RF voltages here and the RF voltages is there but it’s only going to be accurate if the models are accurate and this is something rob touched on in this presentation your design is only going to be as good as your models another important note good thermal design is critical in all i sees but it’s very critical in gun people tend to overlook it now we’ve got a high power density four millimeters for words per millimeter that also means we’ve got a high level of power dissipated in a small area now silicon carbide the substrate for this process that’s got a great thermal conductivity it’s got a thermal conductivity better than gold you’d have thought it however this is an important point that’s at room temperature the thermal conductivity of gold varies very little from room temperature and 200 degrees centigrade the thermal conductivity of silicon carbide drops by a factor of 2.5 now if you look at data sheets for commercial devices certain good commercial devices you’ll see that it has thermal impedance data which varies with ambient temperature that’s because of this effect and it’s very important to consider and understand when your design in a guy I’m not try and I see you as the designer a responsible for ensuring that the thermal performance is adequate and if you get it wrong then your reliability will suffer so the preliminary design we started out with some simulations and we looked at a one millimeter transistor for by 250 micron initially we biased out 100 milliamps that’s a 100 milliamps per millimeter obviously and this is the classic jean max curve we see sure you’re all familiar with this this point here is where we transition from a region of conditional stability here to a region of unconditional stability here so round about our band here the device is unconditionally stable which is nice and we’ve got a gain of 14 DB so we believe when we account for the matching networks of bias circuitry the game flattening the low frequency stabilization got a lot of gain down here we have to do something to make sure that that gave one causes problems with all that taken into account we should be able to achieve 11 to 12 DB again per stage so we move forward with our design we start looking at this transistor initially we’ll look at it in a 50 ohm system and we see what we need to do in order to conjugate Lee match it to 50 ohms we’re not calling we’re not going to conjugate Lee match it I will say that but we start off looking or do will conjugate Lee match it I with a point three nano henry inductor that pretty much takes us to the 50 on point this is one of dem nitrides high operating voltage advantages the output impedance for an out certain output power is higher this makes it easier to achieve broadband operation and even if you don’t want broadband operation it makes a whole

match in process simpler we then undertake some load-pull simulations and we look at the load impedance we need for three different cases we’ve got here this is for the saturated output power and that’s optimum piece at Point and we get 4 watts of output power 46.8 percent efficiency and 43 dbm output IP sorry for the 4 watts of output power this second point you see the optimal match pointers move slightly this is matched for pae we get forty six point eight percent and this final plot here we’re looking at the match for best IP three there we get 43.1 dbm at this particular bias point and these contours here show what happens as we drop the power level DB 4d be at the input so there’s a little bit more work involved and then you end up with this I I want to present some details of the complex modulation and the effects of the pier or not so I’m not going to huge amounts of detail in the design but I’ve got a nice clear plot of the layout this isn’t what they give you on data sheets these days town was they used to give you a nice clear image of the IC and you could see what was going on now they blur it all but this is a nice clear image you can see we’ve got two output devices one millimeter reach their combined here at the output this is some capacitive matching see we’ve got two capacitors here we’ve got two capacitors there because that means we can use a larger capacitor and smaller capacitors of degraded ESD performance and the logical pasta is also less moremore immune to variation with process tolerance we’ve got an inductor here which is through which we inject in our gate bias these transistors can draw a reasonable amount gate current so we can’t have to hire gate resistance it’s a modest gate resistance here it’s a gate bias point here gate price point here and moral capacitive decoupling here Oh what we tend to do this is for the lower frequency d couplet and what we have is a resistor to a capacitor and they all aim of this is just to kill that gain at lower frequencies now the neicy shouldn’t require microwave DC microwave decoupling off-chip you’ve done something a bit wrong if it does all your microwave decoupling all your RF decoupling that should be on it’s your low frequency decoupling you want off check we’ve got some our CD cup in here and what we do is we tend to fill in the spaces like we’ve done here with this to put in as much capacitance as we can without excessively increasing the chip area we’ve got a driver stage here which it’s a little bit less than two-to-one ratio you have to be careful with a gun IC of having two smaller driver stage it’s a classic mistake people make we’ve been used to design in p.m. Pisces and they work out the size of the driver they should use on the tional metrics but the end up with something too small because gallium nitride is a very soft compression characteristic so you’ll end up with a multi-stage I see that you need to drive 10 DB f into saturation in order to compress it so make sure you don’t go too small in your driver stage we have the driver stage which we split here and we’ve got low-pass LC matching these are capacitors these are buyers these little ellipses and ultimately we need to e.m simulate this it costs a lot to make i sees and people like them to be right without having six or seven runs ideally we like them right on the first time now what I’m showing you here is this is a commercially available gas i see that’s the outline for it that has similar performance in that 15 gig bond actually it’s got a couple of DB less ip3 so we’re got a much smaller IC and we’ve got slightly better IP through 23 DB b 33 which will mean we can push higher order modulation through that PA and achieve adequate modulation fidelity that’s what we’re aiming for this is the simulated performance of of the IC see the gain there we got 22 and a half DB nice broadband gamepop across our band of interest slightly widen our band of interest we’ve got 22 and a half DB of

gain this is the 3d be compressed output point have to be very careful when you’re looking at commercially available gonna i sees this is my top tip via they have this metric piece at watch out for how far into compression you need to drive that I see to get to the piece up point some of them are fine but but some of them you know I’ve seen 10 DB and this is our POA here and that’s round about the thirty-seven percent mark for the two-stage I see this is output ip3 here and we’re achieving 46 dbm at the 22 DB per tonne level so this is the simulated performance summary for Ric in the conventional microwave metrics world so this is what we had as our target specification so we’ve got that band with some guard Bank top and bottom we’ve got a little over 22 DB of gain input return loss is greater than 15 DB a DB sorry i’ll put return loss greater than 14 DB we could have achieved better output return loss is actually quite straightforward but we pay for it in reduced IP 3 r.i.p 3 at this particular bias 46 DBA piece adds great and 38 DB MPAA 36-percent and quite a bit smaller than the commercially available gas parts and it needs to be because the gun parts are more expensive per unit area what we did in X is we started investigating the output ip3 how it very bias now initially we had 100 milliamps per millimeter by us we actually move to 130 milliamps per millimeter by US during the design process because that gave us an IP 3 advantage what you can see here this dark blue trace if we go to 160 million per millimeter we get quite 34 d be more crewmen if we go to 200 milliamps per millimeter bias which we’re not going to do because we don’t believe that is thermally judicious you could see you get some some peak in here again all this stuff is only true if you’ve got good accurate models now fortunately if you go to a good foundry they have spent a lot of time and money generating these models they still require a lot of genius skill to use them effectively and get the most out of them but you can get simulated performance that much as well to what you end up measuring so a few words about higher-order modulation here’s a base station actually some building with a base station on top of it you see these Alan bricks hospital down the road you’re not allowed to use your cell phone inside of it but they’ve seen fit to install tons of these base stations on the top and they have a lot of these line-of-sight lengths and that is doing the backhaul so all your teenagers down in the latest tweets or the gaming apps all of this is to Kate of them and there’s tons of stuff to back off all your iphones continuously pole in the network for your mapping features all polling the network all generated tons of traffic to back off and that is why the amount of data we need to bark collars going up and we’re being pushed up the modulation schemes qty 256 now it’s pretty standard to operate with two to five six eight bits per symbol five twelves being rolled out as is 1024 and in honesty that looks like there’s no end in sight and moving up this modulation scheme places greater demands on the microwave front end in particular power amplifier but also in the oscillators which need very good phase noise here’s some constellation diagrams this is quan 64 you can see the 64 points on the eight by eight this is crime 256 11 to eight and 15 12 and you can see they get closer and closer in the end there’s no space between these points it becomes desperately difficult to make sure you can discern one point from the other now fortunately our friends in the digital world in view introduce some clever coding schemes which mean we can tolerate some bit errors and we can work things

out however you do need to pay a lot of attention to preserve in modulation fidelity here’s some eye diagrams now this is quang 64 now here you can’t see what’s going on here you can’t see what’s going on so what we have I apologize from teach anybody to suck eggs in the transmitter we have a root raised cosine filter through which we pass the data and that helps maintain the modulation spectrum to a nice tight bandwidth we transmit it comes in the receiver we have another route raised cosine filter in the receiver so overall we have a raised cosine filter and what that ensures is at the sampling instance here and only at the sampling instance we have no is I into simple interference everywhere else these will raise raised cosine filters they actually caused in inter symbol interference you can’t see what’s going on now we have here 7i openings and eight sampling instances you need an open eye so you can tell one point from another and you can imagine this is in the eye and in the queue domain will have another eight sampling points here 164 and saw the same with Kwan 11 28 1 2 5 6 and 15 12 and you can see they look remarkably time you know these eyes now what happens as we’ll see when you pass this through a circuit with distortion is the eyes start to close this shows the modulation spectrum it’s the pure modulation spectrum this blue trace the route raised cosine filtering has helped us keep this in a nice tight self-contained spectrum this green trace here is the mask the legislative mask for the class for age transmission and there are various of the features of the transmission 56 mcgoats bandwidth symbol rate that this is for the quantum 56 modulation and we have a data rate of 368 megabits per second we’ve normalized here to a one hurt scale and if you multiply each of these by 46 megahertz you’ll get to the class 4-h transmission mask so as I mentioned earlier the disadvantage the higher-order modulation schemes they’re great because you fit loads of data through a small bandwidth however the corrupted by the nonlinearities of the transmission system the PA is one of the main that one of the major contributors and these nonlinearities cause closure of the eyes and that leads to bit errors because we can’t tell what bit were transmitting what symbol were transmitting and we can look at that in terms of degraded error vector magnitude we see generation of intermodulation products in nearby channels and that gives ride rise to spectral regrowth so we see the sidebands growing and growing and eventually they’ll come up above that spectral mask and you’re obliged to meet that spectral mask so we can predict the extent of these problems that our PA causes by using a radio system simulation so we generate random data streams using a crime 256 modulator we do rope root rose case on filtering and then we apply that signal to the PA and we look at the PA spectrum the eye diagram of the constellation and we calculate the EVM here’s the model that we put together so you don’t run the bitstream developed here quanta 256 modulation all in the digital domain here root raised cosine filtering we have a look here another route raised cosine filter in our receiver and this is what we’re putting in so we can see our eye diagram or constellation on our spectrum and those were all the great uncorrupted plots are showed you earlier we now we’re just the drive level to get to set the level to what we want to go into our PA and then we have a model for our PA and it’s a level dependent model and we have a voltage sorry a magnitude and phase transfer function of that PA which we’ve generated from our circuit simulator in this case we used a jelyn abs no doubt

soon-to-be keysight 80s some gain phase correction another route raised cosine filter and we look at what the output iron constellation the spectrum and the error vector magnitude we look at what they look like they’re what are we actually transmitting after our PA so what we did after implementing this system model was to cross-check what the our p3 output ip3 performance predicted by or a circuit simulator could we reproduce this with our system simulator because we should be able to it’s a simple one just two tones in we can’t do this what chance have we got a predicting the before moments with a 256 clamp modulation unfortunately we could so what we did is we generated the level dependent amplitude and phase voltage transfer characteristics for the PA in 80s we took them over to our simulink simulator and then using that system I told you but with two input tones we generated this maroon trace here and you see it agrees very nicely with this blue trace here which is the output IP three verses level that we generated in 80’s and you know that this is this is not on extreme scale this is quite a small scale this is a very good agreement so that gives us confidence to move forward with our system simulations we adjusted the drive level to the PA in order to give us an output power of 29.4 DBM now we didn’t randomly select 29.4 DBM as some sort of optimum power it corresponds to a 60 B back off from the 1d decompression point we had the PA model with the 1 30 milliamps per millimeter bias and we observed the impact that this P I odd on that constellation plot the eye diagram on the spectrum and we calculated the EVM now the EVM target for 256 we we like 1.25 percent I’m not saying if you go up to 1.5 percent you’ll fail you won’t it’s much higher than that but one point two five percent makes us feel really comfortable that’s a really nice safe margin and this is what we got and we got two point three percent so is it it’s a bit higher than one but things to notice can you see the eyes close it now particularly particularly this one here really looks like someone who’s had too many glasses of wine and stayed up too late the constellation it’s not as neat things are starting to merge and can you see how it’s bent as a classic sign and the spectral regrowth it’s really getting a bit close to us makes it makes it a little bit nervous so what we did now is to move to the 160 million per millimeter bias which we know through our circuit simulations gives us improved ip3 we adjusted the level into our PA so we get an output power 29.4 DBM so a comparing like we’d like exactly the same output power and now what we see the first thing to notice is we’ve got a nice healthy margin on the spectral regrowth area that’s good that makes us feel more comfortable the constellation plot is cleaner you can see gaps between all the constellation points it’s not perfect but it’s never going to be and the eyes you can see clear gaps between all of these sampling instances and the vm is one-point-three percent which is actually very good you can receive quam 256 modulation with a lot worse EVM than that so our simulations 160 million pious we maintain the PM up 29.4 DBM and it reduced the performance degradation compared to the 130 milliamps per millimeter I mentioned these points on the last slide but just to recap the iron constellation much less corrupted we’ve reduced the EVM to one-point-three percent which is pretty close to our target or 1.25 % and this target to be honest is it’s a very safe target and the spectrum is well contained within the mask a 7db margin so a few conclusions time nitride is well suited to the realization of 15 gigas pmma sees for point-to-point link applications now with that quarter micron process in on a stake it’s 18 19 gigahertz you can go up to that if you want to go above that for all the other popular ka-band point-to-point link applications you’ll

need a smaller geometry process but gun itself appears to be very well suit to this we’ve highlighted the key steps and the design process the PA itself from a a conventional microwave design metric as good performance 28 volts bias 130 milliamps per millimeter we’re seeing 22 DB of gain an output power of 30 DBM at 3 DB compression thirty percent thirty-six percent paea and an output ip3 of 46 DBM sorry 46 tbm at 130 millions per millimeter and if we tweaked up the bias to 160 million per millimeter we see that increase in 248 and a half for communications applications we need to operate the PA back door from full compression backed off to a sufficient level there it avoids modulation distortion and we presented some system simulations to show that we can achieve the key performance metrics the adjacent channel power the EVM for the quantum 56 applications we saw that the PA demonstrated a spectral mask that was win the XE limits Annie vm of one-point-two percent and an output power at an ALP apparel 29.4 DBM with the bias 160 milliamps per millimeter I think that’s my last point then you offer us any questions first Thank You Liam the reef enlightening presentation and actually preferred around for my presentation but i have a question did you actually measure EDM and 15 gb hurts or you just simulated know we’ve not made this i see ya so we have made gun i sees right but because we’re a design consultancy if we make something for someone they’re not always happy for us to stand up and tell everybody else what we’ve done because they paid for that privilege so this one’s not yet been made there’s every possibility it may be wait may be manufactured and i shall come back and tell you what we measured that’s tantalizing in your presentation is that you but you seem to be suggesting that for the same power capability IP three gallium nitride actually has a better Deenie arity if I understand it correctly no I’m not quite saying that it’s not so much I mean gallium nitride can be a little bit more efficient but but the key thing is what I’m saying here is that for the same sort of diet for a much lower diarrhea you can get an improvement in ip3 using gone compared to Gus that’s really the message I’m trying to get over because I mean you did mention your hints of the fact that again goes much more gently into compressions right than gas but I just want what there is in the technology allows it to be like that or if that’s not like this or dinner no it’s not at all so the actual fight that gun goes softly to impression can be something of a pain it can catch the on worried person out you need to take that into account you also need a lot more work with your modeling and again i’m going to reemphasize what what Rob said earlier you’ve got to get your modeling right if your models not right you can design yourself a lovely circuit when you make it it’ll be miles away and the convention is there’s a lot of work involved to get the model of the gun transistor right because that’s soft code compression characteristic is a little bit different from all the conventional models that were originally developed for based technologies and PMPs technologies which have a much sharper characteristic but if you let’s say you let’s take the leap of faith that the model does work well then you can use gallium nitride to develop a dye which has improved linearity and a much smaller diarrhea than an incumbent Gaspar what’s the reason they’ve got is much more gentle it going that does to do with the semiconductor physics so I I don’t want to I don’t want you it sounds like I’m portraying it as an advantage it’s not it’s actually a bit of a disadvantage so it starts compressing and it compresses and compresses and you know to really get it

saturated you might have to compress it 5 6 DB on a single stage with a pea hemp process you’d probably get it into saturation up to 3 dB at the most so that’s not actually an advantage of gun that’s a disadvantage right it’s a feature of the technology I think it’s just about to the previous question is in eventual people DB point is higher so you’re able to back off further so you generate the same no it’s actually the other way around your P 1 DB point is actually a little bit further away from your piece art and this is this is why i was i was paid taking particular point to mention about the size and the devices because if you sighs the devices of a multi stage amplifier incorrectly you’ll make that even worse and your p 1 DB will be further away from your piece art and you know don’t take my word for it look at some of the commercially available decisis you can see they can have p 1 DB 670 be away from the piece art if you imagine you’ve got this soft compression characteristic that goes on for a long time and then you’ve got the same on the input stage but you’ve sized it too small so it starts compressing here and it goes on forever and really you can buy commercial available parts which require you to drive them 10 DB into saturation to move to meet the advertised piece art can you tell me the amount of gain current you were expecting you’ve mentioned that cake as a potential issue with your lair it’s not huge to be able to be honest but as you drive it into compression you do get more a very from process to process and different boundaries specify different amounts of gate current and then change it again when they get onto the release of a new process and in honesty if you’ve got a mature gun process that is well designed it’s not hugely higher than a gust process but when you’re driving it hard you might see maybe even as much as a million per millimeter I’m a local experience of creeper company job they’re not mentioned yeah they’ve improved but God yeah hey I don’t let me tell you what happened it had a lot higher gate leakage than that yeah particularly if you try to use the maximum drain voltage that you had deposited with the hammock training took Kate could cause damage if you weren’t too big on the trip I want to get most out of it I’m not particular company I think we won’t name have will have a second generation of the process which is much improved in that respect but what what it is to some extent you can almost think of it as the onset of breakdown of the onset of Drake drain gate breakdown and you really don’t want the transistors to have too high a leakage current it’s a sign of a good processes the leakage current is nice and low but nevertheless when you’ve got a large amount of gate width on the device like that that output stage could easily when you’re driving it hard tape perhaps 2 milliamps and and if you’ve got that 2 milliamps through a 1 kilo ohm resistor you don’t need to be a mathematical genius to work out there it’s shifting that bias point around quite a lot so that’s why we try and keep it nice and long thanks okay Terry I’m going to mention a specific country Harry 1 over over West called has got stands for three something think you’ll know who that is worth what Triton’s try could have risked quite recently announced hello low-noise amplifiers in it can’t appreciate this is something of a derivative main theme of your talk there but I do have an opinion on the likely to take up of both power and low noise in this table I have an opinion on everything Terry thanks for asking yeah so so what gun isn’t fantastic for noise but it’s not a disaster either so what people do with these are knees and they call them robust Ellen A’s and they’re pretty nice parts those triquint parts so you buy us them out sort of 10 to 15 volts the knee the knee voltage in gun is round about the 5 volt level so you can’t go too long you buy them at 10 15 volts rather than the 28 volts or even 40 volts you would for a PA and the big benefit to them is that they can withstand a lot of input power so you can do either do without a limiter or do with a much lower power limiter which has lower insertion loss so you end up with a smaller system potentially a slightly cheaper system and what you’re trading is a slightly higher noise LNA

for a reduced performance limiter okay thank you you