#566 – Switching Converter Engineering with Carmen Parisi

Chris Gammell: This is the Amp Hour Podcast. Released November 14th, 2021. Episode 566. Switching Converter Engineering with Carmen Parisi. Welcome to the Amp Hour. I'm Chris Gammell of Contextual Electronics.

Carmen Parisi: And I'm Carmen Parisi, Applications Engineer at Renaissance Electronics. Hello, Carmen. How are you? Pretty good, Chris. Feels different to be on this side of the microphone, finally.

Chris Gammell: I was going to say, you should have introduced yourself as former co-host of the Engineering Commons podcast as well.

Carmen Parisi: I actually haven't done that for a long time, so it's been quite a while.

Chris Gammell: It's on the resume somewhere, right? It is, yeah. You can see that down the LinkedIn page. I was going to say you just dug the mic out of somewhere, so you had it. You had it going on.

Carmen Parisi: It made it in the move last year, so I had to power it up and make sure it still worked.

Chris Gammell: Yep, yep. Well, Carmen was one of my first phone of friends when I moved down to the Triangle area, because you are also in the Triangle area, so that's pretty great. Yeah. And we actually got to meet face-to-face. It was, like, weird.

Carmen Parisi: Yeah, it was the first time. I mean, we've been talking online for how long now? Like 10, 11 years? Oh, 10 years at least, yeah.

Chris Gammell: Yeah, totally.

Carmen Parisi: Right about when the Amp Hour started, because then we wrote for engineer blogs together and various other ventures over the years.

Chris Gammell: In that time, you also then got really good at chip stuff, and I kind of leveled off.

Carmen Parisi: Your word's not mine, but yeah, I've picked up a thing or two.

Chris Gammell: Yeah, yeah, yeah. I think we can just focus on your improvement more than my lack of improvement if we want to do that. But yeah, you've been moving up in the ranks, man. You are a senior applications engineer now, so what does that mean? Oh, actually, I'm staff, if we're going to be pedantic about it. Oh, staff. I'm sorry. Yes, I was too far down the LinkedIn page. Sorry, staff applications engineer. My apologies. So what does that mean?

Carmen Parisi: Yeah, so apps engineer means many things depending on what company you go to. But at Renesas, which at the site I work at used to be Intersil before they got bought out, I do everything from helping define new parts, like what features they're going to have, the pin out, things like that, to documenting them, writing the data sheets, helping customers and FAEs get early design ins, validating silicon, debugging it, and then, you know, kicking it out the door to production and starting on the next chip.

Chris Gammell: Yeah, and that's the whole gamut right there. That's pretty great. That's all you need to know. Do you have a staff, though, like a walking stick, and then you can, when the power chips don't make it through the testing, you can be like, you shall not pass.

Carmen Parisi: Just shoot some lightning out of it and ESD zap them? Yeah, exactly. No, no, I think they give that to you on your 15-year anniversary. Ah, okay. All right, you're going to work on that. Yeah, I got to put in some more time. The senior staff, yeah. Yeah. Then they got to enchant it. It's a whole process.

Chris Gammell: Yeah. So I think we should disambiguate as well because so an applications engineer is someone that's basically at the factory, whereas a field applications engineer is someone who might be kind of on the boots on the ground, out talking to customers, like on customer sites. Is that a fair assessment?

Carmen Parisi: Yes. Yeah. So a field engineer, you know, there is some specialization in field engineers, but yeah, in general, a field engineer has to know a little bit about the whole product line that their company makes, at least enough to go find the right people. Whereas I specialize specifically in computing power regulators for Intel, AMD processors, and a few non-Intel and AMD things, but not so much these days. Yeah. So yeah, I'm more of the inch wide, but a mile deep versus mile wide, inch deep. Yeah. That's a good way to say it. Yeah, exactly. So if your FIE can't answer a question, you know, and it's one of my products, they'll come over and ask for help. And yeah, the quote unquote experts, or at least the ones that pretend to be, will help.

Chris Gammell: Yeah, that's cool. And I mean, so what is the difference? I mean, I guess when you have a chip kind of going through this process, you know, it takes a couple of years to get all the way from an idea to something designed into a product, like an end product and like one of your customers' products then, like where do you have kind of parts in the life cycle at a time? Do you just have one thing? So you're only at the beginning of this, this chip's life cycle, and you're just going to talk to customers and figure out what they need? Or do you, are you doing that while you're also testing the last round of power chips?

Carmen Parisi: Yeah. So at any given time, you'll have two or three different, you know, irons in the fire projects going on at various stages of development. So yeah, I'm supporting legacy products as question comes up, as questions come in. Don't have any silicon actually right now that I'm actively debugging, but I have some that's been fabbed out by the designers and I'm expecting back right after Christmas. And then there's some newer projects I'm working on that, yeah, we're trying to nail down the feature set, what the pinout's going to be, try not to get too bad of a PCB board layout, but also meet the design rule check that the transistor designers have to do.

Chris Gammell: Oh yeah. Yeah. Yeah. And so let's talk a little bit about that, that interaction as well, because so then there, that's kind of the other side of the house is now you're talking to people that are actually designing silicon. But I think when you and I met up a couple of weeks ago, you were talking about like, yeah, but those guys don't really get to like try out chips and like be on the bench and like all the fun stuff.

Carmen Parisi: Yeah. It's rare that you see a designer in the lab. Usually they poke their head in for, you know, a week or two when silicon first comes back to make sure nothing's extremely broken. And then they go back to designing the next chip and the next chip, because there's always a backlog of design work. And then you only call them over if something goes wrong. So yeah, originally when I left school, I had taken some, I see design courses. I thought it would be a lot of fun. I read a lot of Jim Williams and Bob Pease that made it sound like I'd be doing, you know, jumping from customer calls to the lab to transistor level design. And that might've been true back in the day, but yeah, it's really hard to wear all those hats at a big company like this. Oh, those guys actually did that. I didn't realize that. I mean, we've talked about them on the show, but I don't think either of them were actually IC designers, maybe Bob Pease at one point.

Chris Gammell: Yeah.

Carmen Parisi: Yeah. You know, it's certainly a lot harder to go to, you know, designing A to Ds to op amps to power and, you know, write app notes about all those different things. I mean, I suppose if you jumped around, you could, but I've been pretty specialized since I left school.

Chris Gammell: Yeah. Just the domain expertise. And so, so you've never, you've never had to do the silicon design and like, we'll talk about doping and using like, what do they call this fab packages, the PDKs or whatever?

Carmen Parisi: Yeah. The PDKs. No. So yeah. And it depends, you know, cause apps engineers kind of, you know, like every engineer, you kind of specialize and find what you're good at. Some apps engineers, you know, are really good at the marketing side of the job and, you know, maybe transition that way. A lot of apps guys go to design if they really like it. I found myself kind of gravitating more towards an architecture level role where I kind of straddle apps and design. And I'm not the one actually slinging transistors and running SIMs at the PDK. I, I worked with the designers to say, you know, Hey, through this, you know, voltage regulator loop, we need to have an error amplifier gain of, you know, XDB and Y bandwidth. And then they'll come back and say, you know, no, given our schedule. Yeah. The best we can do is, you know, this and that, and can you live with it? And that's my job to figure out like, can we still hit, you know, the specs we're supposed to hit the regulation specs with, with that sort of thing. So I understand to a certain degree, what sort of topologies they're using for amplifiers and different biasing schemes. And I've picked up a thing or two, but yeah, I don't actually have to do the actual design. Yeah. Yeah. I mean,

Chris Gammell: I'm sure it is a lot of tweaking these days. Like I, well, maybe that's, maybe that's a bad assumption, but like power regulators don't seem like they would change that much. But then again, when you were telling me about some of this stuff, I was just like, Whoa, these specs on these, especially because we'll get into a little bit later, but like the fact that you're powering motherboard level processors with like just huge cranking amounts of current, it's just like, Oh my God, this is, these are levels of current that I just, I never, ever experience or think

Carmen Parisi: about. Yeah. So when a lot of people think of, you know, voltage regulators and stuff, they might think like they're 7805 linear regulator, or, you know, you just go on eBay or Amazon or something and you buy a switcher module that you just slapped down and you don't really think about the design. And those are relatively simple. You know, they still have their own complexities and gotchas and whatnot, but when you're doing like processor power, it's, it's a lot more advanced. Actually the, I see itself as like a whole mixed signal PCB, just in one package. You have sensitive analog lines, digital buses, and then depending on the level of integration, you know, power transistors all in this tiny little, you know, four millimeter by four millimeter package or something. And you do get to see the whole system. Yeah. Yeah. It's like comparing like a

Chris Gammell: 1980 Honda Accord versus like a McLaren F1 or something like that. You know, it's just, both are cars technically and very different in terms of specs and performance output. So yeah, that is, that is a pretty stark difference I'm guessing. Yeah, for sure. So what, what are some of like the, give us like an example thing that, uh, so like an Intel processor, like we talking like 20 amps, 50 amps, what are we talking here?

Carmen Parisi: Uh, so it varies, you know, each processor vendor has a wide range and it depends on what it's going into. Uh, over the years I've worked in chips that have gone in cell phones and I've also worked on, you know, big, super high power, multi-phase regulators that have gone into supercomputers. So I've kind of spanned the whole gamut there, but currently I'm doing like mobile chips. So, um, a few of my designs are going to go into more like tablet, ultra book type things. And those processors will run at, you know, 15, 20, uh, Watts thermal design power, TDP. And then some of the higher end desktops that are borderline or higher end laptops that are borderline desktops might pull, you know, 45, 65, a hundred Watts, maybe, uh, TDP. And the bulk of that is driven by the core rail that our regulators power. So even on one of those small 15, 20 watt processors, you could still be pulling peak currents of 150 amps. Yeah. Very, very briefly because even at a volt, 150 amps, 150 Watts is much bigger than 15. So it's a very brief time. So you have lots of really extreme transients. Not, not much, uh, not much curve to be had there. Yeah. Yeah. So it's a lot of extreme transients and then you're also trying to maximize battery life. So the processor is putting itself to sleep a lot. Oh yeah. Yeah. If you were to get your oscilloscope out and look at, you know, uh, it's usually called like the core rail or, you know, something along those lines for the processor. You look at the core rail and your oscilloscope, it's not going to be sitting there at a nice static one volt. Uh, as you run various programs and workloads, the processor commands it dynamically on the fly to different voltages. So you can range anywhere from, you know, say two volts down to, you know, 0.3 volts and then sleep states where you're doing zero current, no voltage, you're just sitting there idle and then you have to wake up and provide 150 amps, 10 microseconds

Chris Gammell: later. Good. Yeah. Yeah. That's pretty, that's pretty nuts. I guess I should be kind when I'm upset with my computer, but still it should be working every time. Come on. I know. Right. How hard could it be? It's just a piece of sand that we've tricked into thinking rock. We've tricked into

Carmen Parisi: thinking whatever the phrase is. Oh man. Yeah. No, we do the best we can to validate them and run them through as many use cases, but it's really hard, you know? And so when you're doing like an Intel chip, I'll just use them as an example. Uh, you have to buy an Intel test tool that they provide to, you know, plug into your board through an Intel provided module that mimics the processor impedances and you pull your loads through that and you talk over Intel's proprietary bus that goes from the processor to our regulator and does all sorts of commands and power state information to and from, and you run it through all the test suites and then you run your own validation on the chip, not just Intel's. And even still, once you actually have a computer running live with software and, you know, various levels of programming and coder knowledge that put the processor through different work modes, you just can't cover every possible scenario. So whenever chips are released, especially for new processor generations that are a different architecture, there's always going to be bugs and updates to the spec and something you got to fix. Oh man. Yeah. That's crazy. I mean, that's,

Chris Gammell: and it's like, yeah. And like you said, I mean, they have, I was just, I was actually mentioning that last week at one thing I was mentioning that, that yeah, I'm looking at a new laptop and it's like a core I 11, I seven gen 11 or whatever the Intel latest are. And it's just like, holy crap. I just keep, keep making new ones. So yeah, it's the computing need is still there, I guess. Oh yeah. And growing every day. Yeah. It's bonkers. So I mean like how, what are the limits going to be? I mean, is it just going to keep getting more spiky? Like we're talking about like the, that super crazy currents for just short spans of time?

Carmen Parisi: Uh, as far as I know it, yes, but I'm, I'm far from a computer architect. So sure. Sure. Someone could invent a new way of processing that completely eliminates that and we could power off a 7805, but, uh, yeah, yeah, no, it's just getting more and more dynamic. And then it pushes constraints on the power designer and the IC companies to provide these tight tolerance windows. Cause as they pull more current, the, uh, the voltage window gets tighter and tighter. You have to have different protection features. It just really gets more and more complicated as the days go on. And it's trickling down to FPGAs, non Intel AMD processors. Yeah. Yeah.

Chris Gammell: Yeah. I mean, the various server processors amounts of rails. They have crazy rails. I guess they've gotten a little simpler than they used to be, but you know, people are using them in similarly crazy

Carmen Parisi: ways. It seems like. Oh yeah. Yeah. If you get one of those FPGAs that costs a couple grand, you're, you're pulling same amounts of current 250, 300 amps, depending on the application. Yeah. It's not the little SparkFun dev board that you buy. That's right. Yeah. Might need a, yeah. 2.5 watt USB.

Chris Gammell: DA. Uh-huh. The good, the good old days. I'm sure that's what you guys say. Huh? Wow. Yeah. So when you are drawing these, I mean, like when you're drawing these kinds of crazy currents, does it require a different like topology than, than that, you know, 12 volt to 3.3 volt buck converter I buy on a, uh, you know, on eBay or something, but like, is it still a state? Actually, I'm staring right at one. Is it still a, you know, they buck topology or is it other like more rare, not rare. What's the

Carmen Parisi: word? Uh, unique. Nope. Yeah. Uh, no. So it's still the buck topology for the core rails and the processor rails. And again, it depends exactly on what, what kind of computer you're, you're working with. Sometimes you're powered directly off the battery. So, um, you know, you could be seeing, sitting at five volts in, and then, uh, if it's a single cell battery, or if you're at three or four cells, you could be at 20, 22 volts in. So depending on the regulator, you get a wider range in down to that core rail of about a volt, give or take. But again, it's super dynamic. Yeah. So it's the buck topology. And then unless it's a really low power, you're going to go multi-phase buck. Uh, so if you think of just a standard buck regulator, you know, you get your upper and lower fat, you drive an inductor, then there's an output capacitor, uh, a multi-phase buck. You have another set, you know, up to as many phases as you could possibly imagine. Parallel

Chris Gammell: sets. It's almost like a, like, like a lock system of, uh, of transistors, huh?

Carmen Parisi: Almost. Yeah. So yeah, you have a parallel set of, uh, transistors and inductors and each one of those is called the phase. So you can have a two, three, four, eight phase system pretty easy. Once you get above eight, things start to get, uh, get a little crazy.

Chris Gammell: And so that is that almost like having like, uh, like eight bucks in line at that point?

Carmen Parisi: Yeah. You're kind of paralleling a buck topology. Yeah. And you all share the output capacitors and the input capacitors. And if you work through the math, you get some ripple reduction on the input and the output, which is dependent on how good your layout is. So yeah, you have to go multi-phase buck to provide the power and also just to slew current that fast because, you know, inductors can't change current instantaneously. It's, you know, V equals LDI DT. But if you have multiple bucks in parallel, you get L divided by the number of phases. So you can slew more current quicker. So you can use less output capacitance than if you had just one phase trying to chug at 150, 200 amps. Yeah. That's bonkers. That's bonkers. Yeah. So it's a multi-phase buck. And then the, uh, the control topologies kind of get a little crazy. You know, it's kind of like your standard current mode basically, but there's a lot of bells and whistles in there to make it perform faster, get super accurate current sensing, because you got to know if you're over, you know, if you trip the overcurrent protection, you want to make sure, you know, exactly when you're tripping, you got to shut down fast. Shut her down. Exactly. Yeah. If you send all of a sudden, you know, V into V out on that buck regulator, straight into the processor, it pops, it explodes. That's in someone's lap. So the protection features weigh in very heavily as you're designing the chips and validating them.

Chris Gammell: Yeah. That's interesting. I mean, what are some of the other bells and whistles you mentioned too? Like, is that the, you'd mentioned like communication between the processor and the, and the power converter.

Carmen Parisi: Yep. So just going to two big processors, you know, everyone's going to know Intel AMD, each one of them have their own proprietary bus that goes directly from the processor to the voltage regulator powering, you know, the main rails. So usually it's a core rail, a graphics rail, if there's integrated graphics, and then just like some type of housekeeping rail that does, you know, some IO stuff and a few other random things around the processor. And those three rails will talk to the processor over these proprietary buses, and you provide current information to the processor, occasionally voltage telemetry, temperature telemetry for what your FETs are doing, not just the processor, but either PCB board temp or FET temp. When they release a new SVID or SVI3 spec update, like it's a, it's a textbook basically full of, full of specs you got to meet. And as an apps engineer in this niche area, it's my job to understand all of that.

Chris Gammell: Yeah. You read, you read every page, huh?

Carmen Parisi: You almost have to basically. Yeah. Inevitably some designer, you know, I'll pass it off and the digital designers will look at it and they'll say, you know, they put this block diagram in here for how they wanted a feature implemented. Do we actually have to do that? Cause that takes, you know, floating point math or something. And we don't want to do that. And can we get it this way? And I got to play interpreter for the spec and call up my contact at one of the vendors and, you know, get clarification and same thing on the analog side. Do we have to do it this way? Could, is this spec optional? And it's a lot of give and take.

Chris Gammell: And I was going to say, you're making the design engineers sound like, uh, you know, like they're whiny, but, uh, No, they're, they're engineers like the rest of us. Optimizing. They're optimizing.

Carmen Parisi: Yeah. We're all, we're, you know, the chip has to, you know, it's got to meet all these specs for processors, but it also has to be cheap enough where people can afford it.

Chris Gammell: Right. That's right. And I'm sure they've got, you know, deadlines and other chips to make.

Carmen Parisi: Yeah. Yeah. It's, uh, it's, especially now, I mean, the chip shortage going on, you know, fabs are busy. So, but I haven't heard about this. There was this pandemic that came up last year and some supply chains got disrupted and I don't know, a wizard did it, but yeah.

Chris Gammell: So do you get an opportunity to go to fabs and stuff like that? Or is that kind of just further than you need to do?

Carmen Parisi: Uh, no, I have not had the opportunity so far to go to a fab. That's yeah. I don't really, in a day to day, I wouldn't have anything to intelligent to say. Yeah. It'd be very cool. And I would happily do it. Yeah. I mean, I just imagine it'd be for funds. No, when I, so over the years, I also worked at TI for a brief time, uh, and at their headquarters in Dallas, you can take a self-guided tour as an employee of the fab. And I never did it. All the times I went to Dallas must've been a million times. And you think I would have just found some time to steal away and do that, but I never did. I have a few, I have a regret about that.

Chris Gammell: Yeah. Hmm.

Carmen Parisi: Yeah. No, never a fab, but I've gone to some pretty cool customer sites. Yeah. I've been to a few different Intel sites over the years.

Chris Gammell: Oh, that's cool.

Carmen Parisi: And then a couple other. They have some amazing fabs. Yeah. Yeah. I've always been, I've been to a few on the West coast, never to the fab, but just seeing the design labs and you know, where they validate motherboards and reference designs that they put out to customers. That's pretty cool.

Chris Gammell: Interesting. Yeah. I mean, I think, I think about like people that have been on the show before too. Like we've, you know, we've had like Robert Farinak's been on here and he was kind of like a processor designer. Dave in the old, old days, he kind of did some high end designs that were kind of in that level, but like just modern day processors. I mean, like the pinouts are just insane. And then the power requirements like you're talking about and just like at that level, it's, it's so, so many levels above where I have ever been. It feels like a whole different world basically.

Carmen Parisi: Yeah. And actually that, that leads to an interesting twist. So as an apps engineer, one of my jobs is to develop, you know, the evaluation board that we use for, for validation for the silicon and then, you know, customer sampling and stuff. And we're kind of constrained by, you know, Intel and AMD have all this test hardware you have to put on there and customers expect that because they're building motherboards. They know you have to include that, but even still, you know, for my, one of my regulators is complicated as it is. Once you consider all the signals and you know, you have to have enough copper to handle the heat and, you know, get the current to the processor without too much voltage drop, I could do a very good layout in, you know, six layers, maybe four, if it's not too complicated, but some of these motherboards are 24 layer boards, 10, 10, 12 layers. Typical. Yeah. Maybe eight. If it's like a real budget, low power, you know, Chromebook or something, but yeah, it's gone are the days of four and six

Chris Gammell: layer boards in my day to day. I mean, what was the era that where there were the four layer boards

Carmen Parisi: for a motherboard though? I mean, that was pretty my time. Uh, when I came out of college and started doing this six was kind of typical, uh, and it's ballooned since there, but yeah, so it becomes a driver of, okay, do I put a, you know, 10, 12 layer eval board out there and most of it's dead copper.

Chris Gammell: Ah, interesting. Yeah. Cause I guess they would want to, they want to just basically start from that point and drop it in, you know, open up an Altium file and just kind of put it into theirs. Right.

Carmen Parisi: That's yes. And no, as a starting point, even still our, even our customer demo boards have a lot of extra test hooks that they wouldn't need on an actual motherboard. But if you're trying to

Chris Gammell: evaluate us against a competitor. Yeah. I just think about like the shape of the, uh, you know, like you guys always have the, like the, the power chips always seem to have like the weird copper, because as you like get closer and closer to the actual pit of silicon, it's just like, it gets really space constraints. And then like, that's the thing that I usually think about, like the copper, the copper planes directly underneath the chip is the stuff that I, that I want to copy pretty much because it's like you had mentioned, you know, when you're working on pin out of a chip, you're trying to evaluate what's going to be the most efficient, but also, you know, it's going to have some trade-off somewhere in there. And I just want to follow that exactly. So I'm not like, Oh, I could probably loop this one over here. And you know, then I've got these loopy EMI. Yeah. Yeah. You don't want to run your current sense lines

Carmen Parisi: directly underneath the, uh, the, one of the inductors. Cause that would just be a bad day for everybody. Yeah. So, so customers do follow what they can, but every motherboard is, is extremely different in terms of density, where they're placing it in the actual computers. It's always a treat to actually get, you know, I do schematic checks and layout reviews for all sorts of customers and stuff, but just trying to get a scale of it. And then all of a sudden you, you get it back cause there's a problem, unfortunately. And you're looking at something smaller than a standard paperback, you'd get it at a bookstore and it's got an entire computer on there. You know, it's maybe the size of my raspberry pie for some of these things. And it's got a full Intel processor on there and the hard drive. And how do you put a laptop, you know, uh, eval board out for that? I don't need all those chips and pin it out. Are you saying that it's, it's this size

Chris Gammell: because, uh, because batteries are taking so much of a laptop these days or because of the style of,

Carmen Parisi: of things that it's going into? Uh, yeah. The style of things are going into batteries, taking up space, you know, everything's trying to get thinner and then, you know, trying to just put an eval board out for that because there's so many layers and so many signals I don't have to care about because I'm not doing a whole motherboard. I'm trying to sell the voltage regulator. I got to make that as good as it's possible. And 12 layers of two ounce copper would make our efficiency and our thermals look fantastic, but that's not a realistic scenario. Yeah. Yeah. Yeah. Yeah. So yeah, it's a big trade-off of, are we close enough to the end application to be a good representation of our performance, but do we make the chip look too good too at the same time?

Chris Gammell: Oh, because if they put it into their product, they're going to be like, I can't hit the same specs. And you're like, well, what you should do is space it out a lot.

Carmen Parisi: Yes, exactly. Yeah. Yeah. When something's the size of a matchbook or whatever, bigger and yeah, you're like, Oh, just, just add some more space between these things. Just double the distance. It doesn't work. And then even still like, you know, we get, you know, leg up with this test hardware from Intel we have to put on there and it mimics the impedance of the processor. As you plug in the load tool that they, these companies provide, well provide like they give it to us, you pay a hefty sum. Right. Right. Yeah. It's included. It's included in the very big price. Exactly. Yeah. They give it to you for free as long as you give them a lot of money. Yep. Yeah. It's meant for the power guys like me, but also the motherboard manufacturers too. So it has pinouts for RAM and, you know, going to the SSD or whatever else it needs. So as I'm going through and doing a layout, I only care about the power of the ground pins and the, you know, the bus and the feedback lines to my processor, but I don't have to pin out an entire processor. So again, you're trying to do all these trade-offs of, is it realistic to have, you know, on this 12 layer board, eight layers of copper going to the PCB or the processor? No, it isn't. So we have to put keep outs and try to do

Chris Gammell: the best we can. Got it. Yeah. So what about on the heat side of things? I mean, so I'm guessing this is very, very high efficiency, but even still the amount of power you're pushing through it, it's not small. So like, what are you doing to then, aside from just big copper pours, is there any other things that you're doing, like, like moving away from bonded to more like chip, flip chip style?

Carmen Parisi: It depends. When I was working on server products, you know, there was a lot of flip chip, you know, and copper clips in there for the power stages and the fats, but that adds cost, which server customers are willing to pay for because, you know, one of the server processors runs orders of magnitude more than our voltage regulators. They'll pay that extra buck for reliability and better performance. But in the mobile world, you know, it's, it's a race to the bottom in terms of price. So you still have to go with, you know, bonded, bonded fats. And there's some room for innovation there and, you know, getting the next best thing, but you know, good layout practices, good design practices from the transistor designer. And then it's on me to help push the customer to, you know, good passive components. Like the inductor is huge, especially on these low power designs for a tablet or something. Everybody wants to go thinner and, you know, everyone's like, well, why don't you just put more battery in and make it thicker? I would love it if it went thicker because then they could use taller inductors. Right. As I'm doing paper designs and working with design to get estimates of, you know, what our chip losses are going to be. And, you know, the FETs are going to, you know, the switching losses, you know, one watt, say just pulling the number out of the air, one watt. And the conduction loss is this. Then the control circuitry draws X number of milliamps. And this is our estimate. And then you add in the inductor losses with models from the vendor or measurements you've done in the lab. The inductor could be 40% of the loss. And it's because they're using some, some paper thin inductor to fit these. It's washed it down. It's

Chris Gammell: got like two windings on it. They're like, well, you know, it may have, you know, a milli-ohm or less

Carmen Parisi: a DCR, but the core losses are through the roof. That's right. And you can only switch so fast with silicon. Right. No one's paying for GAN yet in the mobile space in terms of processor power. Got it. So yeah, thicker inductor lowers your core loss and all of a sudden makes it look a lot better. So yeah, when a customer says we have this height requirement, you always got to ask how, how strict is that? And if you get in early enough, you know, you can have a little bit of clout, but yeah, a lot of that decision's out of our hands. You just do the best you can. They say inductor and you say how high. Exactly. Yeah. It's a fun trade-off though. It's a lot of interesting problems to solve and optimize for.

Chris Gammell: Yeah. I guess I don't think about that mobile space as much, but yeah, that is, they're trying to do that too. They're trying to get as much power out of, out of the same kind of spit. Like I think about even just, again, my only reference right now is that I've been shopping for laptops, but you know, like browsing around in the ultra books and the tablet flip flip arounds with the keyboard. And even like the, the, I guess the iPads of the world too. Right. I mean, those are running their own chips these days, but like, or not Intel chips, I suppose, but like, uh, they're just trying to pack it all in there and do it with less. It's mostly battery inside an iPad. Right. It's like, then there's just a single processing board and it's like super tiny and super constrained. And that's what, that's what the people want, but they also want it to, you know, do crazy amounts of processing. So it's, it's a, it's a crazy

Carmen Parisi: trade-off. Yeah. And one of the ways they manage heat from a system motherboard level is, you know, the processor throttles itself. So you do a big burst of activity, you know, there's different power levels. The processor hits and, you know, assuming you're at thermal steady state room temperature, or typically a laptop or something like that runs about 40 to 50 degrees C inside of it is ambient. So you're sitting at ambient temperature, processor will pull max load and the spec on that is, you know, we could pull that 150, 200 amps, even on that mobile processor for a millisecond, give or take. And then we drop down, you know, from whatever the highest power level is power level 10 down to power level nine. And the, you know, the voltage drops a little bit, the frequency that the processor is running at drops a little bit, and maybe you only pull a hundred amps instead of 150. And then you pull that a hundred amps for another millisecond, maybe two. And then you drop down, you know, way down to 20 amps or something. And then you, you repeat the cycle. Hmm. So yeah, it's, it's huge transients all the time commanded to other voltages. So, you know, you pull that high current at a higher processor or sorry, a high current at a higher load where the regulator runs a little bit more efficiently, you could clock faster, get your work done and then go to sleep and sit in a sleep state for, you know, a millisecond and come back out of it.

Chris Gammell: Huh? That's crazy. Yeah. Yeah. What about going in the other direction now? So, I mean, you're talking about the processor, sorry, the processor talking to the, the regulator is the regulator then carrying upstream, caring about what's going on upstream with the battery. Or is there like a controller normally that you're, that's also sitting up there for like charge recharge. Is that separate or is that something that the parts that you work on also do?

Carmen Parisi: A little bit. Yeah. So it depends again on the actual application and servers have telemetry all over the place cause they need a hundred percent uptime and they want to monitor everything. And someone wants to also do it remotely besides in the system, but just kind of like on a typical laptop, there there's a little bit of a talk. Like usually there's some power monitoring where you might feed through what the overall system power is doing from the adapter or the charger to the processor. So the processor knows overall the health of the system. And they, if they see like the input power is, you know, coming from that, that charger is close to its max value that you put in some register value, you know, say you can provide, you know, one of the USB specs, 45 Watts or something, you know, you're sitting at 30 Watts, but the processor knows it's not pulling power. It knows not to all of a sudden go to max load. All of a sudden it'll run at some lower frequency, you know, lower current to not stress the input power too much. So there, there is a little bit of feedback throughout the whole motherboard that goes around, but it's, yeah, it's a really complex

Chris Gammell: system. It's amazing. Anything ever works. Well, I mean, some of that you see, like, I guess I think about it, some of that you actually see at the OS level, like I know windows and well, I guess all the systems do it, but like, I see it very visibly with windows where it's like, oh, you know, you can set when you're on battery power, you can set, you know, what you're, they kind of do these broad sweeping terms of like max, max performance, max battery saving sort of thing. Like that's the span, but that's really just a monitoring thing. I imagine it happens all under the hood, you know, firmware, like low level kind of way, like almost at the BIOS. Like I wouldn't think that it'd be an OS

Carmen Parisi: level thing. Yeah. I imagine it's, it's in that level too, but that's approaching an area of motherboard design. I don't know a ton about, I've just picked up a few things over the years

Chris Gammell: from FAEs and customers. Yeah, no, it's, I mean, it's really interesting. I mean, that's just the stuff again, I don't really think about it. I think sometimes, sometimes I stop and think about the fact, sometimes it hits me rather that the computers that we use every day are just electronics. I know that's a really stupid thing to say, but like, it's so far above the level that I operate at that I just don't consider electronics anymore. It's just a computer, right?

Carmen Parisi: Yeah. It's just this abstract thing that, yeah. Yeah. But no, inside there, yeah. Ohm's law, Kirchhoff's law is still going on and all that buoy and algebra is happening. Max law is still there. Yep. Yeah. And inductance. Yeah. Inductance everywhere.

Chris Gammell: Yeah, totally. I mean, you'd mentioned like some parasitics and stuff too. What are you usually worrying about when you're designing and this sort of thing? I mean, I guess you'd mentioned with the inductor itself, but are there other like parasitics you have to watch about on chip that are problematic?

Carmen Parisi: Yeah, absolutely. So, you know, if you're trying to get, you know, a smaller footprint to fit on these smaller motherboards, maybe instead of having a separate, this is assuming say you have, you know, the full, you have your controller, your gate drivers and your FETs all integrated on a chip.

Chris Gammell: Yeah. Oh, actually. Yeah. Can you walk through those, those, those terminology as well? Cause I feel like especially new people, that's kind of confusing.

Carmen Parisi: Yeah. Yeah. So, all right. Zooming back out for a second, neglecting like the inductor and the output capacitors for a second, there's three main components to one of these buck regulators. There's the controller, there's a gate driver, which drives the power FETs. And then there's the power FETs themselves. And you can have, you know, again, up to eight, 10, 12, 24 phases. If you're on one of these high power motherboards or graphics cards or whatever. And then you can have some level of integration between those three. So you can have a controller chip, a gate driver chip, and then discrete FETs. You can integrate the FETs in the gate driver. And that's called a Dr. Moss for driver Moss combination.

Chris Gammell: Oh, really? I didn't, I didn't know that term.

Carmen Parisi: Yes. Yeah. Dr. Moss. Dr. Moss. Yeah.

Chris Gammell: Dr. Mario.

Carmen Parisi: Yep. That's when the driver and the MOSFET are integrated. You can integrate, you know, the gate driver with the controller, which has some advantages and disadvantages, and you can't really scale that to a bunch of phases. Okay. But it's good for a low phase count design in certain scenarios, or you can integrate all three. And that's called, you know, like a, we would call that a regulator, which has the actual FETs

Chris Gammell: inside of it, as opposed to just a controller. Okay. I usually say this is probably incorrect, but I usually said a controller has everything external to it and a converter has everything internal to it. So that's, you're saying a regulator has everything internal to it.

Carmen Parisi: But again, it's just, you know, terminology different companies use. Yeah. They're, they're both equally valid. Okay. Yeah. So, but if people are on the

Chris Gammell: Rennesa site, that's, that's the kind of terminology to be able to search for and see.

Carmen Parisi: Yeah. Yeah. Or, or you would call it a PMIC with integrated FETs, you know, power management. I see if the FETs are integrated, depending on the marketing guy, it changes terms.

Chris Gammell: And then does the, do the, do they ever, do you ever do like SIP types things where you have them all as separate silicon and, but then bond them onto the same base die?

Carmen Parisi: Uh, I've done some. Substrate, I suppose. Some, some designs where there's like stack silicon or two die that are bonded together, but it's kind of rare. Usually if everything's

Chris Gammell: integrated, it's, it's integrated. Yeah. I just wondered about like the, uh, the, the process, like, I just think about the FETs are just like these big honking beefy things. And then like you had mentioned earlier in the show, like the, I mean, the actual, the controller is like this, it's a tiny microcontroller basically, you know, it's like actually like pretty, pretty fancy stuff in there. So I would imagine the scale difference

Carmen Parisi: is pretty crazy. So the processing, you can, you can integrate everything on one die. Uh, if you want it, typically if you're buying a, a Dr. Moss, a driver Moss from somebody, Dr. Moss, it's a gate driver silicon die. And that's then bonded to the FETs somehow, either through, you know, copper clips or bond wires. Um, that's usually, you know, a gate driver and a separate FET die. But if everything's integrated, obviously it could very well just be on one die, especially if you're buying a chip scale package. Got it. And so inside a driver as well,

Chris Gammell: there is, it's basically like a current driver to drive the gate of these big FETs, like big trench FET style things, or why is there a driver in the first place? Because if you just have some logic

Carmen Parisi: level PWM signal, you are not going to slew the, yeah, the big gate capacitance of these MOSFETs. So yeah, you need a high powered driver. Got it. And then again, inside the driver, it's not just a simple buffer, if you will, you know, there's different protections you can put in there. You know, if it's a driver MOS, you can get what's called a smart power stage and integrate the current sensing that you need to keep the phase currents balanced between your first phase and

Chris Gammell: your 10th phase. Yeah. I mean, these are snapping on and off too. So I imagine that it would be like, you're dumping like pretty, well, I guess it's not a lot of current because it's going into a FET, but, or gate, the gate of a FET, but still, you still drive it with fast. Yeah. Two, two,

Carmen Parisi: three amps worth of current to slew these things. Yeah. Yeah. I mean,

Chris Gammell: you're driving a capacitor basically, right? I mean like a really fancy, fancy capacitor.

Carmen Parisi: Yeah. Yeah, exactly. And you know, you, you, there's pros and cons to go in discreet versus integrated designs. You know, obviously if you go discreet with this silicon shortage, if you can't get FETs from one vendor, you can switch them out for FETs from another vendor, kind of optimize your system. You know, you can buy a driver perfectly suited to the FETs, but then, you know, if you integrate certain things, you know, driver and MOSFETs, pretty good combination. Cause then that FET vendor gets to optimize the driver for those specific FETs, not any specific FET. Yeah. So you can get better efficiency because your driver and your MOSFETs are paired to one another. So you get the best gate drive, the most minimal dead time. I think you guys brought this up on the last amp hour episode that you don't want the high and low side FET to conduct at the right time, because then you have a short from VN to ground that could blow things up and it really hurts your efficiency. It makes the FETs less reliable over time. Really hurts your efficiency because it's dead. Exactly. Yeah. Well, you don't, you can get shoot through and not blow up if it's just a little bit of shoot through. But the problem is, as you do that over time, the FETs break a lot faster. You kill your long-term liability. So one day it will explode. Got it. Yeah. If it's catastrophic shoot through, you'll know right away if it's yeah, just that tiny bit. That's the one that creeps up on you. Got it. So yeah, you, you optimize that dead time where you can minimize it and know you're not going to get shoot through with a driver moss combination paired up. And then obviously you get the best space savings if you put it all in together, but then you're limited by how many phases you can do on one die, what kind of power you can pump through it, things like that. So yeah, one of the things that we're zooming back in, you asked about parasitics and things you worry about. So as you have those different levels of integration, you can only have so many pins on the chip. And if you were doing a controller with driver moss, each driver moss has its own say five volt supply or 3.3 volt supply for that gate driver. And that's coming across the board and hitting each chip. As you combine it all into one, all of a sudden you get one, maybe two input supply pins, not counting your main VIN supply. And that input supply has to drive your digital circuitry, your sensitive analog circuitry, like the band gap references, the current sense amps, you know, that only have so good a PSRR and these big FET gate drivers. Inside the chip, you could have all sorts of switching noise and spikes on that line. And you can't just slap a decoupling cap in the middle of your chip because it's on chip. That takes a lot of dye area. You can only get so good a value. So yeah, it's definitely a tradeoff between how much integration and how good your process is in terms of noise rejection and all sorts of stuff I don't quite understand. But yeah, it's a huge tradeoff from a design side.

Chris Gammell: It's a system level thing. It sounds like for sure. Like it's, it's really interactions between different subsystems on, on the silicon, let alone outside the silicon. Another thing that sometimes just catches me and blows my mind is just like, like looking at, you know, even just a simple dev board that I'm doing. And it's like, you know, Oh, this tiny little, you know, QFN part and everything's like, you know, it's just a little QFN part sitting in a board with like maybe a couple other processes, probably a couple other components sitting around it. And it's like, there's a tiny piece of silicon within that QFN even. And just like the amount of things that we're doing all the way down to get down into that silicon and then get everything back out again. It's just, it's bananas.

Carmen Parisi: Yeah. It's really, depending on your control topology, you know, there's analog control, there's digital control. If you have a digital controller with all the bells and whistles, like you basically have a computer powering your computer. That's, that's what you got basically.

Chris Gammell: Go dog.

Carmen Parisi: It's computers all the way down.

Chris Gammell: That's right. I mean, so, so how would you, what is the difference there? I mean, the analog versus digital is something like monitoring a process and then selectively changing a PWM frequency versus like with the, with a register or with a, with some kind of code versus like a analog feedback, driving a voltage to frequency converter. Like what does that actually look like in the, between the two?

Carmen Parisi: If we start with analog feedback, you know, there's a classic voltage and current mode control where you're, you're monitoring the voltage on the output. And that comes into an error amplifier with your compensation scheme around it. That goes into the PWM modulator itself that then goes to, that spits out a PWM high, low signal that goes to your gate drive that turns on the FETs and, you know, cycle through the output filter process repeats.

Chris Gammell: Right. Around and around it goes.

Carmen Parisi: Yep. In current mode, you can add, uh, instead of just having a ramp generated on chip in current mode, you're, you're feeding current information in to generate that ramp. And there's a million different ways you can get current. You can go peak current, valley current, you know, monitor from the high side, the low side, you can measure, you know, the average current through the inductor. There's all sorts. And then how you measure through the inductor, there's, you know, four or five different ways you could figure that out too. So you get your current information somehow and you feed that back to the controller and that helps, you know, goes into the modulator and has its pros and cons. And we could do a whole, whole show about different control schemes, but it's all analog signals. It's op amps, it's comparators. Yeah. Right. Right. Right. Yeah.

Chris Gammell: Tiny, small signals get, get fed back through. And then the accuracy of your onboard resistive dividers or your external resistor dividers end up impacting how things go. Right.

Carmen Parisi: Yes. Yeah, exactly. So yeah, all analog circuits, you know, there's some digital in there. It's, it's mixed signal, but you know, and then.

Chris Gammell: Yeah. So like a digital in that case might be like a, a bit that turns, you know, that enables a different gate or a different section of the chip, that kind of thing, or.

Carmen Parisi: Yeah. Like fault signals and stuff like that are digital in these things. And then with digital control, you know, as soon as you come in, you still have your analog voltage and current information because you just have to measure. That's what your output is, is a voltage. That's how the world is. Yep. Yep. So you have to measure your voltage and your current, but then when you get on chip, you go immediately into an analog to digital converter and then you process everything in the digital domain. Got it. So you still have digital current mode control, digital voltage mode control, but it's all done digitally. Not necessarily on a processor. It could just be hard logic gates because that's faster than running a whole processor. And you need, like I said, quick response time to these massive load transients that happen. You can get load frequencies, you know, processor, we have to check a processor load frequencies up to one, two, three megahertz. Oh wow. Repetitive load frequencies. Yeah. Uh huh. Uh huh. Highly unlikely it runs that way very long. It's called usually virus mode or something along those lines, but. Virus mode? Yes. Yeah. There's a virus mode where if you get a catastrophic virus, you know, the processor should behave this way. Please account for it in your design. Really? That's really, wow. Huh? Yeah. It's pretty, pretty interesting stuff. Wait, what, why do they, so there's a virus

Chris Gammell: mode because they expect the processor just so it can handle weird transients happening on the processor. So it doesn't blow up. The virus wouldn't be able to destroy a system.

Carmen Parisi: I think something along those lines, I'm sure there's someone who could trace the whole history back, but yeah, like virus mode is like extreme repetitive load transients at a very high frequency. Huh? You know, maybe it's a lot of hackers. Yeah, exactly.

Chris Gammell: Made my job harder. Bastards in your basement with your hoodies and your dark web.

Carmen Parisi: Your elite keyboards and whatnot. That's right. Yeah. Your cherry MX keys and. Exactly. So yeah. So these digital controllers have to have fast digital logic. It can't just be like, you know, Arduino level speeds. So you have hard coded digital logic, but then for a lot of the stuff, there's, there's processors and microcontrollers in there to do housekeeping tasks.

Chris Gammell: Yeah. Well, it's gotta be like super, super pipeline too, right? Cause if you just think about the, the speed you need to react, I mean that doing a conversion is needs to be fast, but then you'll set the process. It's super fast and react. I mean, I know less about digital architecture

Carmen Parisi: than I do analog architecture at the transistor level. Yeah. Well, I think it's a different guest, but yeah, it's, it's incredibly tricky. Yeah. And there's pros and cons to each, uh, each control scheme. I've worked with both over the years and yeah, it's just incredibly complex. Even in the analog regulators, there's still a lot of digital features to talk to these processors cause you have to have the proprietary bus interface. Sure. Sure. You know, I'll host a telemetry and this and that. Yeah. You need like, uh, like registers and stuff like that. I'm sure. Yep. Oh yeah. You have to store a whole bunch of settings and registers. And then, you know, we talked about integrating the controller, the gate driver, the MOSFET, as you pull stuff in and do different levels of integration, you can also integrate, you know, the compensation components, current sense filters, things like that. And you can have, you know, a digital wrapper around your analog core where you're writing a register saying, adjust the compensation gain from, you know, 20 dB to 30 dB. And you're actually changing resistors, capacitors, or, you know, if it's an error amplifier gain, changing in a transistor to provide more current and boost the gain. So even though it's a quote unquote digital interface to change the compensation, you're actually affecting R's and C's and gains. Uh, whereas in all digital control, even if you have, you know, there's no R's and C's to look at, which I found from an analog guy like

Chris Gammell: myself, very hard to wrap your head around. Yeah. It does it, does that mean you switched to doing like from like Bodie plots to like Z transformers and stuff? That's I think totally

Carmen Parisi: wrong, but, uh, not me personally. No, I've stayed mostly in the analog realm. Yeah. Any digital stuff I've helped out on, I've not been that product line expert that the FAE goes to. I can answer basic questions, dive a little deeper. And then I know to go find the actual digital guy to answer the questions as things get really hairy. Yeah. I play, I play mostly in the analog domain.

Chris Gammell: When does that decision get made to make something like, I mean, cause it sounds like it's a different even team. So like, when does the decision get made of like digital versus analog? I would imagine that's pretty early in the process. I mean, control scheme wise, I, I, you know,

Carmen Parisi: how, how a product line spins up as a business decision, you know, whether it's, we'd identified a need and had to home grow it, or we acquired a company that made digital controllers that, that could fold them in. And then do you cancel the whole product line while they're selling? Well,

Chris Gammell: they can do all this. That's, that's something for the suits to decide. Sure. Sure. I just mean, I don't really understand the, aside from like the, what you were saying with like the speed benefits, like analog seems like it kind of inherently has some speed benefits because there's less complexity, but there must be some benefit to doing digital. Is it like when things are really crazy in terms of like control needs and like being able to like super dial stuff in? Like what is, what is that?

Carmen Parisi: Well, one of the big benefits of digital is, you know, same at doing the PC board level. Like, why would you slap a micro on there? Well, you get firmware, you can do changes without, you know, spinning a whole nother, spinning a whole nother chip. So with the digital, there's always some level of low level firmware that you're writing for these regulators. And if there's a processor or, you know, some issue comes up during validation or when the product's out in the field, you could find a workaround and do a, a, a firmware patch that you give the customer and they can rework the boards.

Chris Gammell: Ah, yeah. Okay. Interesting.

Carmen Parisi: So you get that. Yeah. They don't have to re-solder anything. They just have to, you know, plug in. If they, you know, have something, some, somewhere they can access the part and reprogram it. That's all they have to do.

Chris Gammell: So like tune, yeah. Like tuning. Okay. Like performance tuning.

Carmen Parisi: Exactly. Yeah. With an analog regulator and external parts, we have to change compensation. Someone's got to get out the soldering iron and change a million motherboards.

Chris Gammell: Yeah. Right.

Carmen Parisi: Like the digital part, you just run them back through the automated firmware and it's still a task and no one likes doing it, but it's simpler, easier. You don't have to buy new parts.

Chris Gammell: Yeah. I'd imagine there's even some way to do it over those proprietary buses that you talked about, that they have some way to push firmware equivalent over that or something.

Carmen Parisi: Not over those buses. Cause that's strictly for processor operation and stuff. So usually there's like I squared C or yeah. PM bus. There's a lot, a lot of these parts nowadays and you can push firmware over a PM bus because it's, it's basically I squared C. So yeah, you can just write the

Chris Gammell: firmware that way. Okay. What does PM bus stand for? It's just like an SM bus, but PM bus.

Carmen Parisi: Yeah. It's like a subset of the system management bus. They're, they're almost one in the same. If you're compliant with one, you're most likely compliant with the other as far as I understand it. Don't quote me on that. Okay. There's probably some subtle differences, but yeah. PM bus, SM bus are almost interchangeable and yeah, it's I squared C at the heart of it. Basically, you know, it's open drain clock data lines. And then there's a dedicated alert interrupt pin that you can use for fault toggling and stuff. And then PM bus sort of fixes some of the registers. I don't remember what they all are off the top of my head, but like say register zero B is always output voltage telemetry. So if you read register zero B, you'll know what the output voltage is at any given time. So it just sort of standardizes that stuff. And then it says, you know, registers two to two F are vendor specific. Look at their data sheet. That could be anything.

Chris Gammell: Got it. Yeah. You know, stuff like that. Huh? That's cool. That's really cool. Yeah.

Carmen Parisi: Yeah. But yeah, in terms of when do you go analog and digital and stuff on some of the parts I work on day to day and help define and get in their low level. It's a ton of factors. Yeah. You know, what kind of timeline are we on? How close are we to tape out? What IP exists already? You know, you could be like, Oh yeah, when we built this amplifier or whatever it is, analog to digital converter, we have all these hooks in there to change how it operates. So sure. If we need to change that, we can just do this, that, and the other thing, it's not a big deal. Some designer spends an afternoon and you're good to go. If it's a new feature, that's when you got to figure out, okay, what's the best way to do this. And sometimes there's a very simple digital workaround where everyone goes, why, why spend time redesigning the wheel when you flip a few bits and you know, the digital guys do their RTL and we get a different code to this digital to analog converter. And it just spits out a different reference voltage or something.

Chris Gammell: That's really cool.

Carmen Parisi: So yeah, there's all sorts of trade-offs and yeah, it's not just one designer, you know, sitting at his computer doing this. There's a team of analog guys, a team of digital guys, there's verification, layout guys, everybody's involved.

Chris Gammell: Huh. Yeah. I'd imagine that, you know, there is interest in trying out new things too. At like, well, you'd mentioned like architecture and like, and like going into architecture with your career, what does that, what does that end up looking like? Is it like trying out new, new topologies or kind of new structures internally? Like you're talking about kind of like those, it sounds a little like tools in the toolbox kind of thing.

Carmen Parisi: Yeah. Yeah. So in architecture, the way I see it myself, you know, if I one day reach that quote unquote dream job, you know, you straddle the design world and the apps world where you're working at the PCB and the silicon level. So it could be something like, oh, we're using voltage mode control now, but it doesn't work. We have to come up with a new control scheme. And what's out there? Who, what does TI use? What does Renaissance use? What does analog devices use? What are the pros and cons? Buy a bunch of regulators and test them and come up with an idea, you know, study the math books and the textbooks and get back and find a new control scheme or a new way of doing things. Or what if we sensed current here or there somewhere else, come up with a, you know, a proposal, work with designers to set the specs, you know, design the part, get it back, test those new features, things like that. So that that's one level. It could also be like, okay, the buck topology is taking us as far as it can go. We're going back to linear, going back to linear regulators. Or, or, you know, what if we went to a multi-level buck or a, you know, some LLC zero voltage switcher or something, zero current switcher, you know? Okay. What does that play into it from a cost standpoint? Now you have to buy extra magnetics or extra transistors. There's, you know, there's a, that right half plane zero, like a boost converter, you know, this topology has that too, which is going to wreak havoc on our bandwidth. So maybe we can't use that right away and solving those big, big challenges, not just like, okay, we have this one chip, we're going to spin it off into version B and, you know, we can hand that to a more junior engineer because it's not a lot of forward thinking, you know, right. It's like turn the crank kind of thing and make sure it's right. Yeah. It's more turning the crank. And I've done plenty of that work over the years as I've worked my way up and, you know, that's where I see myself going is these big architecture feature level changes. I mean, do you, uh, you're also writing app notes. What does that look like? Uh, so unfortunately I don't know if I'll ever be like a Jim Williams that like writes, you know, uh, what's the one app note that's like the, you know, a hundred page app note on,

Chris Gammell: uh, Oh, high speed compared. No, it's like app note 45 or something. Yeah. There's a third. I always think of the thermocouple one. That was always one of my favorites. That one's good too,

Carmen Parisi: but this isn't the one I'm thinking of and now it's going to bug me, but I'll figure it out later.

Chris Gammell: You know, you are in front of a, one of those magical things called computers, so you can look

Carmen Parisi: it up if you really wanted, but, uh, and then people got to hear me type and my mechanical keys

Chris Gammell: won't be the right. Oh, I see. We got one here. He's a hacker folks. No, no, no, no. I have the

Carmen Parisi: cheapest Logitech Amazon had for my computer here. I spent it all on RGB.

Chris Gammell: All my budget. Got it. Got it. So what, what is that process like? I mean, like you basically, you're just, it's, it's almost like writing a paper it seems like, or what does that look like?

Carmen Parisi: Yeah. So, uh, I have put out a few app notes. You know, I had a couple articles early on for, uh, like EE web. I posted one about voltage rating. Uh, I came up with a neat little circuit. Uh, so day in, day out, I designed buck regulators and optimize the hell out of them. But as you do validation and whatnot, you get all these little one-off scenarios where there's just no good solution. You can't just go to, you know, tectronics or keysight or whoever, and get a piece of equipment. So I kind of came up with a neat little circuit to do fast input voltage transients that can slew hundreds of microfarads of input capacitance. And some of these boards that I work on and check the line transient response of our regulators. So I came up with a neat little circuit and that got published in EDN seven, eight years ago, something like that.

Chris Gammell: Yeah. Back when it, back when it was actually a decent magazine.

Carmen Parisi: I, I guess I kind of proved your point. I was going to say, I don't know. I haven't read it in a while.

Chris Gammell: I miss, I miss when it was better. I don't think, I think it was, you know, we were not there in the golden days, but, uh, I just miss, I miss the technical magazines

Carmen Parisi: being decent. Yeah. Well, uh, these days I've been reading signal integrity journal and that's a, that's a pretty good one. Yeah. That's a pretty good site. I recommend some of it goes right over my head. Cause it's, you know, 56 gig, you know, Pam for modulation, you know, equalization techniques or whatever, but there's a lot of good power integrity articles. Steve Sandler writes for them. So does, uh, Eric Bogutin, uh, Isvan Novak, a lot of these power guys that you'll see like design con and around the internet, right for signal integrity journal. So it's a good place to look for for articles and whatnot. Okay. But anyways, yeah. So a few of my app notes and whatnot that I've written kind of came from ideas in the lab and things that I've noticed during validation and whatnot. And then two of them I put out were just quick little checklists, kind of repackaging stuff that's already in the data sheet, but in a more digestible form for like, okay, this data sheet's gigantic. Like just give me, give me the highlights. So that was the purpose of those, as opposed to being a end all be all textbook. When I was at TI, I did put out one app notes, uh, SLVA 882, I think is the official number. Multi-phase buck design from start to finish part one. Yes. Yeah. That was easier to

Chris Gammell: Google Google for than the others. I'm still struggling. Oh, I think they're all on my LinkedIn.

Carmen Parisi: If you want to ever link them. Yeah. Yeah. I'm pretty easy to find on LinkedIn and you can, if anyone's interested in reading ancient articles, yeah, they can, they can pull them up and I can tell you all about the things that had changed about them now, 10 years later. That's right. Isn't that fun with

Chris Gammell: the internet, uh, keeping, keeping all our history? Yes. Yes. It's beautiful. It makes writing a resume

Carmen Parisi: easier. Yeah, sure. Yeah. Yeah. But, um, yeah, uh, the multi-phase one was, uh, something that came out from my time at TI. I was kind of in a unique position. I wasn't doing traditional product line applications work. I was working for, um, a marketing guy. I don't think George listens to this, but if he does shout out to George. Hi George. Who, who took me on. He was a marketing guy that was just generating so much work that the regular product line couldn't keep up with him. So he got his permission to hire an apps guy to work directly for him. And George and I went after a lot of non Intel business in the server space for providing power. And that was a fantastic, fantastic role. And so I was doing a lot of education because, you know, we talked about FPGAs pulling 250 amps or whatever. Well, a lot of people doing these network interface cards or hardware accelerator cards. We're used to using, you know, the low end FPGA that pulls an amp or two and, you know, you plop down your whatever jelly bean switcher and it works just fine. And all of a sudden they go up the next rung of the ladder and it pulls only 30 to 50 amps or something. And now they need a multi-phase regulator and it becomes a whole thing. And so I was doing a lot of education and, you know, what is multi-phase? Why do you care? Where is it good? So I pitched that idea to my boss and said, like, you know, Hey, I put together this whole presentation. I gave a million times over the years. I wrote that app note. I turned it into a three-part series that's on TI's YouTube channel. So that was a lot of work. That was a lot of fun. That app notes gotten some traction. Actually, I had an old boss, not George, a different boss at TI recently texted me the other day and say he was going through the YouTube rabbit hole and stumbled on some professor somewhere that was referencing my app note in his lecture. So that was pretty cool. That's nice. Yeah. That's great. Yeah. So that was pretty neat. Yeah. So I originally pitched that app note series as, you know, a three-part series as, you know, what is multi-phase, how to do layout for it, and then, you know, how to do compensation for it, you know, with TI's parts, obviously. And then bit off more than I could chew and everything picked up and I was doing too much other work. And then the years went by and all of a sudden I never wrote it and I took another job. So my apologies to everybody looking for part two and three.

Chris Gammell: Yeah. I'll do that someday. Someday. Yeah. Yeah. Just for funsies though. Not for, not for TI. Come on.

Carmen Parisi: Not for TI. No. Maybe for Renaissance one day I'll write part two and then maybe someday I'll go somewhere else and write part three and I'll just be, yeah, big, big company spanning guy with this trilogy. Yeah. Yeah. No. So that app note came about from wanting to teach and, you know, I saw a need for this new generation of people who not necessarily are power experts or processor experts or whatever, all of a sudden needing to care about processor power. So that's where I came out with that app note. And then for TI's power supply design seminar, I had a paper accepted on designing for high slew rate, high load step transients, which got accepted and got published last year, actually, in 2020. Oh, cool. So you can find that. Yeah. That one you have to make an account for. It's not just free, but you can just make a free TI account and see that on their site. And what is it called again? If you go to ti.com slash PSDS, here I am working for Renaissance and pitching a TI link, but whatever. I'm pitching myself really. I mean, the seminar we've talked about on the show before. So yeah. Yeah. The power supply design seminar. It's pretty,

Chris Gammell: pretty well known in the industry. You're not, you're not blowing any minds here. Although if people haven't heard about it, I will tell, I will say that it's, it's pretty great. It's unfortunate that it's been, just took a break because it's COVID, but yeah. And yeah, I was supposed to go travel,

Carmen Parisi: present my paper, but yeah, COVID happened and I ended up taking this Renaissance job. So yeah,

Chris Gammell: right now. Do you guys have similar kind of like public facing resources for Renaissance or, I mean, is Renaissance more targeted towards kind of like B2B customers type stuff?

Carmen Parisi: I imagine there's something I can't say for better or worse that I've spent a ton of time digging through Renaissance's site and looking for a big application note dump. Okay. I just happened to know where mine was on the TI website. So yeah, if you go to ti.com slash PSDS and you go to 2020 resources, you'll see, yeah. Voltage regulator design and optimization for high current fast load rate load transients. And if that doesn't put you to sleep, you can read my paper. And that one's only two years old now at this point, if you count from when I wrote it, but that was as I was dipping my toes into power integrity and the stuff that Steve Sandler and Eric Bogutin talk about all the time. I was really starting to wrap my head around that. And I wrote that paper as kind of, you know, trying to put my spin on it. And now I would change a million things about that paper, but I still stand by it. It was a lot of work and I'm proud of it. That's great. That's great. Yeah. So that was the last one I put out. And yeah, most of the time though, I, you know, I don't have a 150, 200 publications that some of these guys do, but you do a lot over email. I've written some, what could be app notes that I'm going to one customer to solve a specific problem and you have to dive deep into the theory, but also to their, their board layout and how one thing affects the other. And yeah, that comes up on the day to day, or especially when you're sitting in design reviews or brainstorm sessions for a chip and the whiteboard comes flying and you delve deep and then you got to write it up and you have internal spec documents for these chips as they get designed. So everyone can stay on track. And most of that eventually spins off into the data sheet, but there are books that have been written by many people that are internal company documents that will never, ever see the last day.

Chris Gammell: Yeah, that's too bad. I've seen it. That they can't like harvest it and genericize some of that, that knowledge that's been formulated in email.

Carmen Parisi: That would be fantastic, but it's so specific to, you know, the company IP, the process nodes you're on, the internal architecture that gives you your secret sauce that, yeah, these people, there was an architecture guide for a generation of chips. I was working on a TI and this thing was legitimately a textbook. It was 200 some pages long written by, you know, four or five of the lead, you know, like designers, apps, engineers, architects, and it covered anything you could possibly want to know on some of these chips, like why they chose this topology for that, for the current sensing. And it was really cool to look at, but it's nothing you could ever publish directly. Right, right, right. Yeah. Maybe someone got an IEEE paper or something out of it. I don't know about it.

Chris Gammell: 4,000 users typing on a forum, huh? What?

Carmen Parisi: Exactly, yeah. So yeah, you can write another app note. Everyone can pump one out on doing type two voltage compensation, but, you know, the number of people that would actually read that sort of thing is the people in our product line that had to. Yeah, right, right. Yeah, it already hit its wide audience. Yeah, so that sort of stuff I do, a lot of technical writing. One thing I've been getting into is learning Sigridi, which is this like board simulation tool, you know, 3D field solver and stuff that Cadence has that bolts onto their board layout software, and it's incredibly powerful and does so many things. So I've been starting to use that to do parasitic extraction. Hopefully soon I'll be correlating it to measurements when I get some boards back. Yeah, that'd be cool. And yeah, using that to drive design decisions so I can show them like, hey, I know that pinout looks terrible.

Speaker ?: I told you.

Carmen Parisi: Yeah, I know that looks terrible because I've been doing this for a long time, but I can quantify it now and look, you know, I got to put the decoupling cap here and that adds, you know, another half a nano Henry of inductance to the path and, you know, that's going to kill it and we might as well not have that cap. So let's move those pins or really try to if we can.

Chris Gammell: Oh, really? So you're using it almost as like a way to kind of preempt arguments.

Carmen Parisi: Yeah. So, you know, as I was...

Chris Gammell: I mean, simulate in the lab so that you can prove your point and win arguments.

Carmen Parisi: Yeah. Well, as you're doing these chips, you know, if it's a turn the crank, we're just going to take, you know, this spin, this chip and, you know, cut off two features that aren't needed for budget builds. You know, that's more turn the crank and, you know, you hand that off and whatever. And that's the sort of work I did out of college and as I rose through the ranks. Yeah, to learn stuff and just... Yeah. So that sort of stuff, there's very little say, you know, the designers, you know, it could very well be the same pinout just with internal features turned off to sell it cheaper to customers that are doing, you know, budget PC builds that don't want to pay for all the features or whatever. So the pinout's fixed. It kind of is what it is. Writing the data sheet from existing material and whatnot, you're starting with an eval board. But if you're doing like a big platform part, like you just moved to a new process node or one of the processor vendors really changed their specs and now you got to design, you know, all these new blocks and everything, you're defining the pinout and the feature set for a part that's going to spin off potentially a dozen or so parts from it. So you start early, you get the apps engineers involved and you look at PC board from the beginning. So as soon as we have a concept pinout, now I'm starting to look at the board layout and okay, where do we put the decoupling capacitors? What size? You know, can we get away with 0402s and pack everything close or do we need 0603s, you know, all that sort of thing. There's monstrous 0603s. Yeah. Yeah. On a motherboard, 0603 is huge. I know. I know. Yeah. Gone. When I first started 10, 11 years ago, I could get away with 0805s, maybe a 1206 cap. And now, uh, no, it's all 603s are smaller. Wow. Yep. Occasionally you can still use bulk capacitors, you know, your electrolytic, you know, 10, 11 caps and stuff. Yeah. Polymer or whatever. Those are still in some designs, but yeah, a lot of customers go all ceramic and, uh, that has its own design challenges. But anyways, as you're defining these pinouts, you know, as soon as you have a concept, you got to start with it because it's important. Every pin location matters and, you know, you don't want to put your high current phase node if it's completely integrated next to your voltage sense lines, just because then your inductors right next to your feedback lines and it gets messy and, you know, it's just not good. So you get design involved early. Uh, well, obviously designs involved early, but you get apps involved early, start thinking, thinking about that sort of thing. And you can push back on design and they can push back on apps and say, well, if we do that, then we, we can't package this thing. We have to move to a more expensive package or whatever. Yeah. Yeah. Um, so it's a back and forth between everybody to get a trade off and what's going to work the best for, or at least

Chris Gammell: reasonably well for everyone. Awesome. You know, if you talk to someone younger who's maybe listening and they want to do what you do, what do you, what do you usually tell them in terms of like, uh, how to, how to get in a similar space? Uh, so yeah, the best way if you're interested is get

Carmen Parisi: an internship at school. I mean, uh, all these semiconductor companies are hiring interns. TI takes a million interns every year. It seems when I was there, I was doing some recruiting for them and, you know, they're forever hiring for, uh, new college grads, apps, engineers, and everything. Interns, you name it, they're looking for it. Same at Renaissance. We're always looking for people too. So an internship is definitely the easiest way. You can dip your toes in the water, see if it's for you. At least the groups I've worked with, we've always had our interns and new hires doing legitimate real work from day one. Yeah. Yeah. Yeah. That's pretty great. Uh, but honestly, uh, one of the big things that I found helped me is I've just stayed curious. I've read a lot, you know, everything, like I mentioned earlier, you know, your Jim Williams, your Bob Pease, all that sort of thing. Signal integrity journal. I've got a whole host of app notes and whatnot, and books next to me in my to read pile that I'm may or may not get to one day. And it's power related. It's, you know, analog related. It's RF stuff. I'm, I'm looking at everything because, you know, I, I focus on power. It's a very niche area, but I, I very much care about edge rates and EMI. Not so much that my products specifically have to pass that sort of thing, but I have to be aware of it to know what trade-offs our customers are making. You know, the high-speed digital certes stuff. I don't do that day to day. It's a lot of it is over my head. It's kind of dark arts a little bit to me, but I have to know enough about it again, to see those lines on the customer's layout board, know what he's trying to do and say, okay, well, I can't tell you to put your power plane right through those. I get why you came up with the stack up that you did. And, you know, just understanding the whole picture. Cause you're exposed to quite a bit. I get what you did, but you're wrong. It makes my job harder. So please just move those,

Chris Gammell: move those DDR lines. I knew you were going, I knew you were going five plus gigahertz, but, uh, you know, we've got some current over here.

Carmen Parisi: Yeah. But yeah, as you dip your toes in the water, get exposed to things. Don't think just because you didn't start out that way in the semiconductor world that you can't work your way into it. Some of the best guys that I've, I've worked with started off in other areas. Uh, I was fortunate enough to get semiconductors cause it's what I thought I wanted to do. I thought I wanted to do IC design. So that's what I focused on, but you know, you get a lot of these guys, some of them might work for your, your Dells, your HPs, your whatever. And they decide they like the IC world and they'll come work for you. Uh, cause then you just get a whole host of knowledge that I'm not exposed to because I've never designed a whole motherboard or a laptop and they can, you know, feed that back as an apps engineer to design and say, no, no, no, this is exactly why customers do this. This spec is important. They just bring a different perspective to it. So, uh, it's just another avenue to break in. Yeah. And once you're in the semiconductor industry, you know, you can see there's a lot of moving around. People start as FAEs and move to product line apps as they're exposed to different things. Apps engineers I've seen move to test or product engineering to design. Occasionally a designer will leave design, but usually people don't, at least from what I've seen, you know, up to management, sure, whatever. But I don't think I've ever seen a designer leave from design to go to apps or test. It's always been the other way around. Oh, interesting. Okay. Hmm. Yeah. So that's something we've, you know, as we're an hour in here that we've barely touched on is all the various roles of the semiconductor industry, uh, chip design company. Yeah. Well, we can always

Chris Gammell: talk about it in the future, you know, have you back. That's true. Yeah. Yeah. Just to highlight

Carmen Parisi: real quick, uh, so people know what I'm talking about. So we mentioned the FAE, there's the apps engineer, there's the design engineer, but there's also a very important role as the test engineer. And when Silicon comes back, I validate it at the board level. They validate it more at the IC level and the wafer level. So, whereas I care about like, okay, how efficiently can I make this buck earlier? How efficiently can I make this buck regulator switch? They'll care about, okay, how accurate is the band gap on a chip and they'll, they'll dial it in. So it's exactly 1.2 volts and the temperature coefficient is just perfect. Uh, you know, Oh, there's a reference current of 10 microamps here. Uh, how accurate is that 10 microamps? Because design told me the whole chip uses this darn thing. So it better be 10 microamps on the dot and they'll dial it in. And a good test engineer is worth their weight in gold. A few years ago, I was working on a part. It wasn't a core power part. It kind of did all the other little peripheral rails that a processor needs. And it had this shutdown spec that it had to meet of, you know, say, I forget what it was, 15 milliwatts or something. And that was all, you know, barely switching light loads. You had to cover the, uh, inductor loss and everything as well. And that, and our chip itself was pulling too much power. And it was a test engineer that figured out why not design. He was looking through the schematics and looking through his test program. And he was like, wait a minute, why are you guys doing, you know, whatever it was, some bias scheme on the chip? Like you have this whole path to ground here. That's burning microamps that no one needs. Uh, what if you just did it this way instead? And design was like, holy crap, that's a really good idea. We'll change that in the next revision and saved us, you know, five microamps or something and lowered our shutdown current even more. That's awesome. Yeah. Yeah. So it's just a role. Not a lot of people know about, but it's, it's a pretty cool role. They do, they do a lot of cool stuff and they're also involved in the definition process too, because design will say, well, we can do that spec that I tell them to hit. Uh, but in order to do that, we have to design X, Y, Z and a test guy will raise his hand and go, cool, but that would require me to do this on my tester program. And how do we do that? Uh, what hooks would you need there for me to do that? And then, you know, again, cause if you design a chip and no one can test it, you'd have no idea if it works. Right. Yeah. Hmm. So yeah, it's very interesting. And then they'll go into on the backend after a part's been released to production and they'll, um, they'll do like this backend cleanup to make more money on the part and get better ROI. Uh, after you've run millions and millions of units and shipped them and whatnot, you can go back and look at the test program that trims the parts. And you can say, what can we do to speed up the test program? Okay. Well, these programs here, we've shipped, you know, a hundred million units of this regulator. We've never, ever, ever failed this test. It seems like it's not needed. Delete it. We've saved, you know, half a, half a millisecond on the test time or something. Uh, boom, you can process wafers faster, more money in our pocket. It's like interesting little backend cleanup or, Hey, what if we adjusted the limits on this test? Because according to a bunch of statistical analysis I did, we can, boom, we can now pass more parts. Uh, again, more good wafers or more good units out of the wafer, uh, better ROI. So it's a, it's a real interesting role that doesn't get hyped a lot because how would you teach that at school? I guess, but it's a, it's a pretty cool job that can add a lot of value to a chip. Yeah, that's awesome. Yeah. Carbon, where do people,

Chris Gammell: uh, where do people find you online if they want to talk to you more?

Carmen Parisi: Uh, geez, these days I'm, I'm pretty lame. I'm not in a lot of places. Uh, the best place would be to just look for me on LinkedIn, uh, LinkedIn.com. I think it's slash car par. I managed a good, uh, good URL, something along those lines. Yeah. Uh, so find me on there, look me up. I'm more than happy to connect with the answer. Any questions you have, uh, not on Twitter anymore these days, so can't find me there, unfortunately. Oh, that's okay. You're not, you're not missing too much.

Chris Gammell: That's what I hear. Awesome. Well, uh, Carmen, thank you for being on the show and, uh, you know, glad we're crossing paths now in person and it's not just online and thanks for telling us about regulators. All right. Take it easy, Chris. Thanks. Yep. Bye.

Speaker ?: Bye.