CB FI 421

Circuit Break Podcast #421

Tracing a Path for PCB Design Automation with Sergiy Nesterenko

Related Topics
Food Device Contest Wrap up

We celebrate the innovation and creativity of Food Device Design Derby entries like the JavAqua, Pizza Pouch, and the winner, BarBuddy.

MacroFab Platform Updates from Kyle McLeod and Nicholas Lundgaard

The episode provides insights into MacroFab's efforts to make PCB manufacturing more accessible and efficient for their customers.

Dr. Duncan Haldane from JITX on Automating Circuit Design

Dr. Haldane on his background, the problems JITX is trying to resolve, & new auto-router plans. What's the deal with the "hyper-aggressive pogo-stick"?

Other Resources

Circuit Break Podcast
Webinars
Videos
Tour MacroFab's ITAR-Compliant Facility

March 12, 2024, Episode #421

Sergiy Nestorenko, founder of Quilter and former SpaceX engineer, discusses revolutionizing PCB design automation. He shares his journey from aerospace to starting Quilter, aiming to transform PCB design into a streamlined, AI-driven process. We delve into the technical hurdles, the fusion of engineering and advanced software, and the vision behind making circuit board design more efficient and intuitive. Sergiy also addresses the potential educational impacts and the future of hardware engineering in an AI-augmented world. Join us for a dive into the evolving landscape of PCB design and engineering innovation.

🚨Contest Announcement 🚨: Introducing a new Circuit Break contest! This contest is themed around building food-related electronic projects. We’re offering over $5,000 in cash prizes, themed trophies, and free prototyping from MacroFab. The deadline to submit is March 31st, 2024. Thanks to Mouser Electronics for sponsoring the contest prizes!

Discussion Highlights

  • Transition to Quilter: Sergi discusses his first experience with PCB layout at SpaceX and the realization that led him to question and eventually automate the process.
  • Autorouters and their Limitations: Exploration of the limitations of current autorouting solutions in PCB design software and why they fail.
  • Quilter's Vision: Sergi outlines the long-term goal for Quilter to become the compiler for hardware, enabling engineers to focus on design creativity rather than manual layout tasks.
  • Organic and Unconventional PCB Designs: Discussion on how AI and automation could lead to more optimal yet unconventional PCB designs, moving beyond traditional shapes and layouts.
  • Aesthetics of Earliest PCBs: Questioning the assumptions of traces leads to a discussion about the designs of the earliest PCBs in history.
  • Simulation Integration and FCC Compliance: Sergi emphasizes the importance of incorporating comprehensive simulations, including electromagnetic and thermal, into the PCB design process for ensuring compliance and performance.
  • Feedback and Learning from AI: The discussion reveals how AI, like Quilter, can introduce designers to considerations they hadn't encountered, fostering a learning environment while automating tedious tasks.
  • Exploring Design Variations: Sergi envisions Quilter enabling engineers to explore thousands of design variations, including different stack-ups and materials, to optimize board designs beyond traditional constraints, enhancing innovation and performance.
  • Neural Network Integration: The conversation delves into how Quilter utilizes neural networks not for direct layout generation but for guiding classical algorithms in decision-making.
  • Feedback Mechanisms: Quilter encourages user feedback through various channels to refine its algorithms and user experience.
  • Future of Hardware Engineering: Reflecting on the future, Sergi and the hosts discuss the potential shift in hardware engineering roles with increased AI integration.

Relevant Links

Sergiy Nestorenko, founder of Quilter and former SpaceX engineer

Sergiy Nestorenko, founder of Quilter and former SpaceX engineer

About the Hosts

Parker Dillmann
  Parker Dillmann

Parker is an Electrical Engineer with backgrounds in Embedded System Design and Digital Signal Processing. He got his start in 2005 by hacking Nintendo consoles into portable gaming units. The following year he designed and produced an Atari 2600 video mod to allow the Atari to display a crisp, RF fuzz free picture on newer TVs. Over a thousand Atari video mods where produced by Parker from 2006 to 2011 and the mod is still made by other enthusiasts in the Atari community.

In 2006, Parker enrolled at The University of Texas at Austin as a Petroleum Engineer. After realizing electronics was his passion he switched majors in 2007 to Electrical and Computer Engineering. Following his previous background in making the Atari 2600 video mod, Parker decided to take more board layout classes and circuit design classes. Other areas of study include robotics, microcontroller theory and design, FPGA development with VHDL and Verilog, and image and signal processing with DSPs. In 2010, Parker won a Ti sponsored Launchpad programming and design contest that was held by the IEEE CS chapter at the University. Parker graduated with a BS in Electrical and Computer Engineering in the Spring of 2012.

In the Summer of 2012, Parker was hired on as an Electrical Engineer at Dynamic Perception to design and prototype new electronic products. Here, Parker learned about full product development cycles and honed his board layout skills. Seeing the difficulties in managing operations and FCC/CE compliance testing, Parker thought there had to be a better way for small electronic companies to get their product out in customer's hands.

Parker also runs the blog, longhornengineer.com, where he posts his personal projects, technical guides, and appnotes about board layout design and components.

Stephen Kraig
  Stephen Kraig

Stephen Kraig is a component engineer working in the aerospace industry. He has applied his electrical engineering knowledge in a variety of contexts previously, including oil and gas, contract manufacturing, audio electronic repair, and synthesizer design. A graduate of Texas A&M, Stephen has lived his adult life in the Houston, TX, and Denver, CO, areas.

Stephen has never said no to a project. From building guitar amps (starting when he was 17) to designing and building his own CNC table to fine-tuning the mineral composition of the water he uses to brew beer, he thrives on testing, experimentation, and problem-solving. Tune into the podcast to learn more about the wacky stuff Stephen gets up to.

Transcript

Parker Dillmann
Welcome to circuit break from MacroFab, a weekly show about all things engineering, DIY projects, manufacturing, industry news and automating PCB design. Where your hosts, electrical engineers, Parker Dillmann. And Stephen Kraig. This is episode

Sergiy Nestorenko
421.

Parker Dillmann
Circuit breaker from Macrofab. Hey, circuit breakers. We have an announcement. We're running an electronic design contest on our community forums. The theme is food devices.

Parker Dillmann
Go to forum.macfed.com to find out more information about the contest and how to enter. There's a category on the left side of the form that says contests, click that, you'll find it. For prizes, there's over $5,000 in cash and free prototyping services through Macfab. And the most important thing, a trophy to show that your design was one of the best entered. There'll be a link in the show notes where to find more information about this contest and how to enter.

Parker Dillmann
And thank you, Mouser Electronics, for sponsoring the contest.

Stephen Kraig
This episode, our guest is Sergey Nestarenko. Sergey is the founder of Quilter, a startup focused on automating PCB layout. Before Quilter, Sergey was a senior avionics engineer at SpaceX focused on the Falcon 9 Falcon Heavy second stage in high radiation environments. Sergei holds a triple major in mathematics, physics, and chemistry from UC Berkeley.

Parker Dillmann
Welcome to the podcast, Sergei.

Sergiy Nestorenko
Parker, Stephen, thank you guys so much for having me.

Parker Dillmann
We were talking a little bit before the podcast about you were working on radiation protection and components for a rocket, and now you're doing basically automated PCB design. How'd you decide you wanted to go do that? Like, did you do layouts before SpaceX or what's the process there? The history.

Sergiy Nestorenko
Yeah. Yeah. It's, there's a few elements that went into this. So my first PCB or my first actual layout was at SpaceX. And I built a lot of electronics before, but I, you know, use breadboards or perfboards or whatever else and just hand soldered stuff.

Sergiy Nestorenko
And, one of my first tasks at SpaceX was to I was actually still in the EMI group. This was before I came over to radiation effects. It was to break out a soft switch, a soft start switch from the Falcon 9 power system, and hook it up to some automated test equipment. And that was my first time opening up Altium and trying to figure out how to do a PCB and how to do layout. And as I was doing it, I asked myself, why isn't a computer doing this?

Sergiy Nestorenko
And then I clicked the button, and I saw why a computer wasn't doing this. And proceeded to do it manually. And I proceeded to build, you know, more than a 100 boards for both SpaceX and personal projects after that. So, you know, later after I left SpaceX, I got an opportunity to start a company, you know, thought about different problems I might solve, and eventually started thinking, what is something I know is a real problem that I saw at SpaceX that I think I can tackle? And this was it.

Stephen Kraig
You know, you said you opened up Altium and clicked a button, and there was sort of a moment the way you presented it in terms of, oh, this is why a computer doesn't do it. I'm curious if you could go a little bit further into that. Like, what about it was that the computer wasn't doing what you wanted it to do?

Sergiy Nestorenko
Yeah. Totally. So what we're talking about for the Avid listener will be auto routers. Right? That's kind of the standard thing that people think about when they're thinking about automated PCB layout.

Sergiy Nestorenko
And there's a bunch of things that went wrong for me and go wrong in general with these things. So the first is just quite literally that they don't handle the whole process. Right? So when you think about a layout, you're really starting with, you know, schematic and some constraints. What is the size of your board?

Sergiy Nestorenko
Where do you want connectors or buttons or monitors or whatever else? And you're going all the way through fab files. Like, how do we actually go and get this manufactured and get this board on my desk? Autorouter is typically just considered with connected dots. Right?

Sergiy Nestorenko
It's not including, you know, placements. It's not including stack up. It's not including kind of high level planning of, you know, mixed signal and whatever else. And then, of course, importantly, it's not including the considerations of simulations and did we actually get it right. So the main things that bothered me are, it's not tackling the whole process end to end.

Sergiy Nestorenko
Even when it is for most boards of any complexity, you don't get a full solution. Right? Autoradors don't typically get to a 100%. And you're kind of left cleaning up the mess and trying to connect the rest. And then the most importantly, and the thing that I think is the worst about Autoroutters today is that you have to review their work.

Sergiy Nestorenko
You have to look at all the traces and ask yourself, is this actually going to function? Are we gonna overheat stuff? Are we gonna have crosstalk issues? Are we gonna be susceptible to radio disability or emissions and so on and so forth? And that can be as difficult or more difficult than just doing it yourself and doing it right.

Sergiy Nestorenko
And that's why, in my experience, they're just not used practically speaking.

Stephen Kraig
When you went to make your first board, was that your expectation that you jump in and press a button and it kinda the the the hamster starts going?

Sergiy Nestorenko
You know, it wasn't, but simply because I was on a floor with 600 other electrical engineers, and I could see their monitors. I could see that people were designing boards. I could see that this was happening on almost every computer around the house. And so I knew that this is a manual task. But as I started doing it, I just thought, well, hold on.

Sergiy Nestorenko
Like, it it would seem like a computer should be good at this. So why isn't it? And so my expectation came from, like, the first principles of of doing, you know, placement, doing routing, and asking, like, this seems automatable. But, no, my expectation given that everybody else is doing it manually was that there's no easy answer to this, and I was gonna have to do it manually.

Parker Dillmann
Yeah. It's one of those, like, computers are really good at pathfinding. And ideally, routing is pathfinding for the for current oh, the traces at least. So if you're just doing that part, yeah, you you would expect that a computer would be at least kinda decent at it, but, yeah, no. It's not really good at it.

Sergiy Nestorenko
Actually, this is one of the misconceptions that I certainly had before I started this problem, and I find a lot of people had. You know, when I explain this problem of routing a board to somebody who's never done it before, I certainly make this analogy of, like, okay. You have this maze and you're you're finding a path from a to b. And, typically, that specific problem is, you know, computer science 1 a. Right?

Sergiy Nestorenko
That's a pretty basic algorithm. You do a grid or run Dijkstra or whatever. Not so hard. The hard part is that once you have a solution to that first maze, it becomes an obstacle to your next maze. And once you've done that 10000 times on your 10000 and first trace, how did you make sure that you have space for it?

Sergiy Nestorenko
And that, I think, is something that humans are exceptionally good at because we've evolved in a three-dimensional world, and we have capabilities to arrange things three dimensionally and then plan and so on and so forth. And that's actually deceptively hard for a computer. That problem explodes so quickly that you can't brute force it. And coming up with heuristics for it, well, that's what autoradors have done today, and it doesn't work that well. So even forgetting the important stuff, even forgetting the physics.

Sergiy Nestorenko
Right? Connect the dots on a dense 10,000 pin board is is still hard for a computer. And it's because of that kind of compounding problem of your next maze solution has been blocked by all the previous solutions.

Parker Dillmann
Yeah. I I think the last time I actually used an autorouter was caught a while ago. But it would try a solution and then, like, well, didn't finish, so I'm just gonna start over. And just I guess it just tweaked the parameter and just ran it again. And I think the biggest thing was you can just set, like, how many iterations it could run, and I think I just set it up on the lab computer and just let it run over the weekend, which probably didn't make the TA really happy, but whatever.

Sergiy Nestorenko
Did it finish after the whole weekend?

Parker Dillmann
No. It got to 94% routed.

Sergiy Nestorenko
Right. Yeah. It's an exponential difficulty increase. Right? So making a router that gets to 50, 60, 70% is really easy.

Sergiy Nestorenko
Getting to 95% is a little bit harder. But going from 99.9 to a100, that is the difference. Right? That is the thing that's actually really, really difficult. And it's no surprise.

Sergiy Nestorenko
Right? If you I'm guessing I don't know for sure, Parker, but I'm guessing that you looked at that 94% completed board, control a, delete, and restart.

Parker Dillmann
Yeah. Yeah. I started all over, and I just did it myself. Right.

Sergiy Nestorenko
Right. But if you hadn't, and I've done this experiment, if you hadn't and you tried to finish that last 6%, you would have been shoving around deleting, replacing, and redoing probably 20, 30, 40 percent of the traces that the auto router did, which is why people don't do this.

Stephen Kraig
Yeah. I've I've told my boss before on a relatively simple board. I'll say, it's gonna take me 3 days. The first day is get to 80%, the next day is to get to 98, and then the last day is 4 traces. And and that's then you'll have a board.

Sergiy Nestorenko
Oh, spot on. Spot on.

Parker Dillmann
So, yeah, we know, like, engineers just don't like autoroutters because we have terrible experiences with them. And a lot of times they're just viewed as wastes of time. So where does quilter? How does quilter? How does it go from the 98 to a 100%?

Sergiy Nestorenko
Yeah. So I think it's, I actually think that that's not sufficient even. So I find it helpful to talk about Quilter first from, like, a super long distant future of the dream of technology and then work backwards about, like, how do we get there? And the thing that I find helpful is to to draw inspiration from from the software side of things. So when I work on Quilter, one of the analogies I tend to make, and I wanna present this just because I come back to it all the time and I wanna put some footing for it, I I think of what we do as kind of the compiler for circuit boards.

Sergiy Nestorenko
So roughly speaking, you can kind of argue that a schematic is a high level human readable representation of logic. Right? Like, what are your inputs, outputs, and functionality that you want from the board? And to a software engineer, that's a lot like writing Python or c plus plus or something like that. Right?

Sergiy Nestorenko
It's a human readable language that's defining your your functionality and constraints. But, of course, in software, we can click a compile button, and in hardware, we can't. In hardware, we have to take that logic that's in the schematic and make it real by hand. Right? So I think about the inspiration for Quilter for, you know, let's call it the 100 year future.

Sergiy Nestorenko
Right? When computers and AI are so incredible that this is obviously gonna be automated. I think about it as the compiler for hardware. Right? Like, what happens when electrical engineers can focus just on schematics and high level constraints and put their creativity into that piece?

Sergiy Nestorenko
And then, like, layouts are so good from computers that it's nonsensical to even check them. Right? Like, in software, we don't look at the assembly that comes out. We don't look at the ones and zeros that are going into the registers of the processor. It's just gonna work.

Sergiy Nestorenko
It's gonna be fine. Wasn't always that way. In the early days, autocode was equally regarded as something terrible, as something that's never gonna work, as something that's inefficient. Like, the same stuff that we talk about with autoroutters today. But today, like, could we imagine an operating system or a web browser or the Internet or neural networks without a compiler?

Sergiy Nestorenko
Like, that would be so inconceivable to do that all from assembly. So, like, the inspiration for what Quilter should be and what I wanted to create in this long distant future is the compiler for hardware. Like, what happens when we have that equivalent of the compiler that we can run and we lower the barrier to entry for electrical engineers significantly. So that, you know, like a 10 year old can go play with Python and make a website today. A 10 year old's probably not doing a PCB.

Sergiy Nestorenko
But when it gets that easy, like, what all are we gonna build? So when I think about it from that perspective, like, it's not sufficient to just get to a 100%. We really have to get to the point where it would be nonsensical to review the output of the layout, the same way it is with a compiler. And that op you have to get to a 100%. No doubt.

Sergiy Nestorenko
Like, if you didn't connect everything, the board's obviously not gonna work. Like, what's the point? But you also have to make sure that the layout is as aware or more aware of the physics than humans. That it's factored in all the electromagnetics, all the thermodynamics, all the manufacturing processes, at least as well and ideally much better than a human can. And that's when we get to the kind of compiler vision of what it can be.

Sergiy Nestorenko
And then, of course, the hard part is it's easier said than done. Like, how do we actually get there?

Stephen Kraig
And that's what Quilter's trying to march towards?

Sergiy Nestorenko
Bingo. That's exactly right.

Stephen Kraig
We've been having some discussions about AI and electronics at work recently. And, just this concept that a PCB, especially just like a very simplistic square or rectangle PCB, is a very conceivable thing for a human. You can look at it and with enough brain power, you can understand it and and dig through it. But it's also really well set up for humans to do that. AI doesn't have to have the constraints necessarily that a human does in terms of visibility or readability.

Stephen Kraig
And so there's this concept we've been talking about of AI generated electronics that could potentially have far more organic shapes to them as it takes into account what you were talking about, like the actual physics behind things. And it may be entirely unintelligible shapes, things that you can't understand by even looking at it, but it actually functions most optimally.

Sergiy Nestorenko
Oh, man. You you touched the hot topic for me. That's very, very, very excited. I'm glad you brought that up. I'll say a couple of things on this.

Sergiy Nestorenko
So one of the most inspiring examples that I saw of this was actually some research done in FPGAs, like, 15 or 17 years ago. There was a lab that was interested in using genetic algorithms to configure FPGAs to basically do And, you know, the genetic algorithm worked. Eventually, FPGAs spit out the right answer. And the humans started looking at, like, well, what did it do? Like, let's go take a part.

Sergiy Nestorenko
How did it actually do the logic here? And they saw that there was, like, a signal chain from the inputs to the outputs. Okay. That's great. They started shutting down the pieces that seemed unconnected.

Sergiy Nestorenko
It stopped working. Then they turned those pieces on, like, wait, it's working again. But, like, they're logically disconnected. Why is that happening? Loaded the program on a different FPGA, didn't work.

Sergiy Nestorenko
So what ended up happening is that genetic algorithm actually found parasitics within that FPGA such that the coupling path across different registers was producing the right answer. That is absolutely insane. So when I think about this you know, I don't know I don't know if you guys have played at all with PCB filters, like trace filters, where you can do, like, high pass, low pass, band pass filters with just trace shapes. Super cool things. We we tested some of our electromagnetic simulations against these.

Sergiy Nestorenko
It makes me wonder, like, maybe there's a case where we make these arbitrary shapes where, like, traces aren't even connected, but the coupling is, like, perfect and does perfect isolation and, like, works out better than a trace. I mean, okay. We're we're really getting into the dreams here. This is not what Quilter is doing today, but, it's potentially very, very exciting. But, like, to bring that more back to earth, one of the first things we did was actually to question the assumption of octalinear traces.

Sergiy Nestorenko
From first principles, you think that a curved trace is typically better from a cyclical integrity perspective, and so on and so forth. And so the boards that we've started generating were any angle traces going kind of naturally anywhere. People hated it. I have had senior engineers at companies look at that and be like, can you even manufacture those? And, like, of course, you can, but it's just so strange compared to what we're used to.

Sergiy Nestorenko
You know, these nice 0 90 45 degree traces that are easy to to shove around and all that stuff that we've actually had to post process those out and, like, turn our kind of any angle representation to something that's more familiar. So there's an interesting discussion here to be had about what's actually optimal from a physics perspective and, like, what are we used to and what are our EDA tools built optimally around.

Stephen Kraig
You know, I find that really fascinating because not only do you have to convince those who are automatically have a bias against the auto router, You have to convince these other long standing biases in electrical engineering specifically, like the angle thing. I can't tell you how many times I've heard acute angles being a problem because of acid traps in the PCB manufacturing. And from my understanding, that hasn't been a problem for decades in Correct. In the PCB fabrication world, but that mentality still sticks around and if you show the right person or perhaps the wrong person something that has that, they'll immediately poo poo it and say this is a terrible design. And so there's a lot of hills to climb across that.

Parker Dillmann
We're talking about the the electrons just can't turn that sharp of a corner, and they just fly off into space. Or or what's really funny about what we're talking about here is so some of the first circuit boards, though, I mean, this is like post wire wrapping, was drawing them on transfer paper

Sergiy Nestorenko
Mhmm.

Parker Dillmann
With markers. Yeah. And they're very if you've seen them, they're very organic shapes, teardrops around where the holes for your components work because that was actually just because people were manually drill pressing through PCB. So hit tolerance wasn't very good, and so you had to have a lot of copper around it.

Sergiy Nestorenko
Yeah. Mhmm.

Parker Dillmann
Oh, even a citation on that, like, we found this, they're chopping up old Cray computers, supercomputers, And so that has some cool pictures of like the chips on them. And so the the reduce the density of those chips, you went from like a 100 mil spacing to 50 mil, but your PCB technology couldn't handle 50 mil hits on for through hole. And so you they had like weird staggered dip packages. I'll show you all the pictures later and post them in the show notes. But it's interesting from, like, they could build these packages, but we didn't have the pcb technology to handle them.

Parker Dillmann
So you they had to play games with the physical construction of how they attached. So, yeah, you have all these, like, organic traces and stuff. And then I think what it is, like, some of the first EDA tools that were in the computer probably only really did 45 degrees was but what you got, and then enough people started using that. And now we have that mentality. It's interesting to go back to how it kinda originally was.

Sergiy Nestorenko
Yeah. Parker, you're completely right. When you do the math of like, one of the common things that you have to do in an EDA tool and for Coulter as well is this notion of, like, point to line segment distance, point to point distance, or line segment to line segment distance. Right? Roughly speaking, you can break the whole problem into those basic operations.

Sergiy Nestorenko
The difference between whether those two line segments are access aligned or or 45 degree mitered allowed, and you consider the cases versus, like, any angle, it's a significantly cheaper calculation if you restrict yourself to, you know, 0 90 degree or even allow 45. So, like, when you have a seventies or eighties level computer, which is, you know, woefully slow and you're trying to do efficient operations, like, yeah, that's a very reasonable thing to do. Today, it makes absolutely no sense. But I think just the history of it is such that that's exactly what happened and we just got stuck with it. And, like, you'll see if you produce, you know, TikTok, Altium, or KiCad.

Sergiy Nestorenko
Right? Like, you can grab a trace and shove it around, and everything else will move with it as long as it's 0 45 or 90. If you have any angle traces and you try to grab 1 and you try to shove it around, you're stuck. It's not going anywhere. And that's a practical pain.

Sergiy Nestorenko
That's a real actual valid reason to kind of be forced to work with those kinds of things. But it shouldn't exist.

Stephen Kraig
You know, I do wonder if these kinds of things are have any kind of real basis or if they're just propagated. In other words, boards that first started using, you know, 90, 45 angles, they just worked, or perhaps they had better performance because they had better constraints or something, and the mentality that that is better without any empirical data just maybe perhaps stuck. I mean, I I do agree that it's easier to route a board with nineties and 40 fives. And so the you know, easy is probably gonna win over optimized in, in in most situations. But when the computer does it, do we really care?

Stephen Kraig
And also, can half of us, if not most of us, back up why a particular thing is better than another? I mean, to be honest, most of our experience with PCBs is we did something and it worked, therefore, we keep doing it. Right?

Sergiy Nestorenko
Yeah. Absolutely. I mean, I think in general, I'd love to see a lot more a lot more, like, math and simulations in in our space. I'll give you a really good example that we struggle with a lot. You know, my 1st year on the electromagnetic interference team at SpaceX, one of the debates you'd have a lot of PCB designers is to split ground planes or not to split ground planes And what to do with ground and, like, dip, you know, star to polish and all this stuff.

Sergiy Nestorenko
And it's like, people just treat it like as black magic. Like, oh, this senior person told me this is the way to do it, and that's the way we're gonna do it. Or I did this and it didn't work in this case. I'm never doing it again. And, like, let's just do that.

Sergiy Nestorenko
But, you know, it's not especially nowadays with today's computers and today's simulation software and, software and, you know, GPUs and all that stuff. Like, you know, running running a full Maxwell equation solver, you know, the one that we use in our stack for a full PCB, for a noise victim query or even noise many victims query, is a few minutes. I mean, it's it's still annoying. Like, the tooling is just such that it's annoying to set up and, you know, has to be made better and automated and all that stuff. But, like, the Maxwell equations have been known for a 150 years.

Sergiy Nestorenko
We've had provably convergent numerical solutions to them for a half decade. We've had fast enough computers to do it in minutes for, you know, at least a decade. Let's start simulating and measuring these debates, not, you know, using best practices.

Parker Dillmann
Yeah. Part of your DRC stack should be the FCC rules.

Stephen Kraig
Yeah. That that'd be really cool.

Sergiy Nestorenko
A 100%. I mean, I'd love I this is not something we're seriously considering right now. But, like, in the long term future of Quilter, like, shoot, I'd love to run an FCC in my lab. Like, hey. We're gonna build the board, and what we're gonna guarantee is that you're gonna pass FCC.

Sergiy Nestorenko
And, like, we'll test it and certify it. And if it doesn't pass, that's on us. Like, we screwed up the simulation. We screwed up the design. We'll go back, fix our engine.

Sergiy Nestorenko
That's really what we care about here. Like, in a compiler world, like, where you don't have to review the design, that's what should happen. Like, the board on your desk should already be stamped for FCC approval at that point.

Stephen Kraig
I mean, that just sounds like you're trying to simulate everything. You're just trying to simulate reality of the board. And, that just that seems like a pretty daunting task. I mean, what's your confidence in that?

Sergiy Nestorenko
Yeah. Well, let's take it apart. So, of course, like, when you think a priori, like, let's take the whole board and simulate it. Like, that's a lot. But what are we actually talking about simulating here?

Sergiy Nestorenko
Right? We are focused on the PCB. So what is the purpose or job of a PCB? Like, the purpose of a PCB is to faithfully recreate the intent of the schematic. And in general, the only thing that the PCB can do is either do that or fail to do that.

Sergiy Nestorenko
You know, introduce some sort of parasitic or something like that that that screws it up. So okay. So then what we're really reduced with is that we have this schematic. We're assuming it's correct. Right?

Sergiy Nestorenko
Making a valid schematic is a different problem altogether, just like writing valid Python code. It's a different problem altogether to compiling it. So let's assume the schematic is logically correct. Now it's a matter of how do we make a board that faithfully recreates it. So the nice thing is that the PCB is a passive system.

Sergiy Nestorenko
Right? So that already reduces us to this kind of set of simulations that is just like, what's basically the coupling between different geometries and not an active antenna doing something weird or responding to something weird. So the way that we think about this is that we have to understand from the schematic and the data sheet and the user, what is the signal that's propagating on every single net? We have to look at, okay, like, we have a spy bus here. Cool.

Sergiy Nestorenko
We have a USB here. It's a differential pair. Cool. We have a sensitive ADC and it's measuring some very, very, you know, low noise signal. And that's the thing we have to be really careful about.

Sergiy Nestorenko
Cool. Great. Then we have to be aware that, like, hey. We have a switching converter here, and that's gonna be noisy. Alright.

Sergiy Nestorenko
Noted. And then we enumerate all the questions. Right? So okay. Like, one of the questions might be, we have the switching regulator.

Sergiy Nestorenko
It's creating a whole bunch of noise. Certainly, the coupling from that to the ADC is not 0. Right? The the the parasitics are never truly 0. But the real question is just, is it below the requirement to which it functions?

Sergiy Nestorenko
You know? So with the ADC, maybe there's a requirement from, you know, the team that's creating the sensor, or maybe we just want it to be lower than the voltage range of a single bit or something like that. Alright. That gives us a level. The crosstalk has to be below x, you know, minus 30 dB or something like that.

Sergiy Nestorenko
So I think that it's it's very possible if you know all of the different signal requirements and all the different nets throughout your board to then enumerate all of those potential concerns of, like, you know, I have a a noise source here. I have a bunch of sensitive things here. Let's calculate the crosstalk. Let's see if it's within spec. Check.

Sergiy Nestorenko
Check. Check. Check. Check. And, yeah, that's gonna be 1,000 or tens of thousands of checks, for sure.

Sergiy Nestorenko
But we do it. It's possible. We do it as humans, and then we iterate until we've solved every one of them. That's kind of how we think about this. I mean, I'll I'll frame this more in an electromagnetics perspective.

Sergiy Nestorenko
There's obviously similar considerations for thermodynamics, but that's kind of how we approach it. We're not simulating what's happening inside the chip. We're taking for granted that, like, there is a certain signal starting at the pin. That's what's going into the PCB. The PCB is a passive element of copper and fiberglass.

Sergiy Nestorenko
We can do the, the simulation equivalent of taking a vector network analyzer to every port and, like, seeing what the s parameters are and what the transfer parameters are, and then see if we've met all the constraints. And I think what would be really cool is at the end of that, let's prove our work. Let me give you, as the designer, here are the s parameters you can plug into SPICE to see how your circuit's gonna perform. And then if you want to challenge it, go hook it up to a VNA and see that you got the same kind of transfer characteristics that we did.

Parker Dillmann
Yeah. It it's kinda interesting to talk about that way because when we all electrical engineers, when we design and do the layout, we are taking an account of all those things even though we're not really thinking about it. And part of that of taking those account might be what we you just talked about earlier, whereas a rule of thumb that someone learned or, like, a senior person has told you to do it this way. Even though you're not even applying, you know, these signal characters, you you are applying a rule that takes care of that supposedly. Mhmm.

Parker Dillmann
So it's one of those you have to start knowing more simulating, so to speak, more about your circuit than you probably initially knew about. Totally.

Sergiy Nestorenko
I think that burden should be on us. Right? Like, I think that as the user of the compiler, you should be responsible for, you know, giving us a schematic. We actually work on taking all the data sheets that came along from your schematic, downloading them, processing them, extracting information about what's on the different pins, checking it with the user. Right?

Sergiy Nestorenko
Like, you should make sure that we've correctly understood what all your circuit is doing. But then, like, why make you create that laundry list of concerns? Like, we should just show it to you and say, hey. Like, we've thought of all of these different things. If we miss something, okay, let us know.

Sergiy Nestorenko
But, hopefully, we haven't. I think that's where we have to get to to make this kind of thing work.

Stephen Kraig
It seems like you have to have, as the designer with this, you have to have far more intimate connection with your schematic than than maybe we've previously asked for. Because what was going through my mind when we're talking about this is coming right out of school, I wouldn't have known half of this stuff, but I was making boards. Right? And a good chunk of what I was doing was a little bit of guesstimation or just like here's what I've learned from school so I will apply that, But I wouldn't necessarily know what parameters to play with or what knobs I have to be able to turn in order to get exactly what I'm going for. So it seems like with this, there is perhaps a little bit of a higher threshold of what you need to know before getting into it.

Parker Dillmann
Well, it does also sound like Quilter is also supposed to be correct me if I'm gonna be wrong about this, but it's also supposed to be kind of like that senior engineer helping you understand stuff about your board.

Sergiy Nestorenko
Yeah. Or

Parker Dillmann
your not your board, but your schematic.

Sergiy Nestorenko
Yeah. Yeah. I I think, Parker, you're right. Steven, I I think that if you had to like, one of the reasons that autoroutters have failed is that they have a lot of configuration knobs. And the designers of autoroutters have just said, like, if you want this to work, go configure it, and nobody's gonna do that.

Sergiy Nestorenko
Mhmm. Never. Right? You're never gonna sit there and learn all these different knobs and play with them and learn what they do and read everything like that. No.

Sergiy Nestorenko
It has to be easier than that. So I think, Parker, you're right that, like, we should actually be more like the senior engineer telling you about your board. Right? We should A lot of the information is there. Right?

Sergiy Nestorenko
Like, let's say you're giving us, I don't know, STM 32 board. Great. As long as you've given us the part number, we have access to the data sheet. If we have access to the data sheet, we can go and query what all are the pins on the s s m 32 and what are their requirements. And then we'll see, okay, this needs a clock.

Sergiy Nestorenko
This needs a power. This the power expects a bypass cap. This is a spy connection. This is a ethernet connection, you know, differential pairs. So as the junior designer, if you've never heard of a differential pair, you'll be like, okay.

Sergiy Nestorenko
I guess Voltor's doing this differential pair thing. Nice. Great. Go do it. That's the level that we should really be getting to.

Sergiy Nestorenko
Mhmm. No.

Stephen Kraig
I can I can get on board with that? It just it does seem, I I agree. If the software is the gray beard that's distilling information upon you, then thumbs up.

Parker Dillmann
Yeah. It's also one of those it will help you learn and because that's this is what I found out using, like, that's for me at least, that is AI's power is coming up with stuff that you haven't thought about yet and, you know, implementing.

Sergiy Nestorenko
Yeah. Completely. I mean, I think one of the exciting things so for today, right, like, I don't wanna oversell. Right? Like, this is a long term vision.

Sergiy Nestorenko
It's a very difficult thing to build. 60 years of efforts. Nobody's done it. We've set out. We haven't completely solved it for sure.

Sergiy Nestorenko
Shoot. Gonna try. And, you know, we've got some great funding and great team and lots of great users giving us feedback, and we welcome more of all of the above until we eventually get there. But it's a journey. Right?

Sergiy Nestorenko
This is a difficult problem to solve. You know, I think, the thing that we can do well now is in the limit of the easier boards. Right? In the boards that are in the limit of, you know, lower frequency, lower power, where this stuff is less of a concern, we can already save a lot of time. Right?

Sergiy Nestorenko
So designs that take a human about a week to do, Quilter can do in an hour, or maybe a couple of hours. And that's a lot of benefit to, hey. One, you don't have to spend that week. You can go do something more important, like start writing the firmware for your board or start doing some testing or validation or environmental testing or whatever else. And the second thing is that board is on your desk a week sooner, which means any gremlins or issues that are there, you find out about a week sooner and your product gets launched sooner.

Sergiy Nestorenko
That's great. That's a good start, and people are already finding a lot of value in that. However, like, the stuff that excites me for the long term potential is making better boards. Right? So, like, what I think about is how many times when you sit down to do a layout, do you make a lot of conservative assumptions?

Sergiy Nestorenko
Right? If you're doing something like a flight computer or a Starlink antenna or an Apple motherboard, and you know it's gonna take months, you're gonna create a lot of conservative assumptions ahead of time to make sure that you get it right on close to the first try or the first try. So you might make your board a bit bigger than it has to be. You might use a more conservative stack up. You know, signal ground, signal ground, signal ground, just for, like, perfect isolation.

Sergiy Nestorenko
You know, so on and so on and so forth. And what the software can eventually allow us to do is say, hey. Why don't we consider 10,000 different variations? Let's consider a bunch of different stock ups. Let's consider a 100 different manufacturers at the same time.

Sergiy Nestorenko
Let's consider a bunch of different core materials, and let's just see, like, which ones actually work and where the simulations check out. And then as a designer, you're not just, like, getting one board across the finishing line to hopefully for the best. You can do this whole trade study of, like, cost versus performance versus, you know, manufacturability versus yield versus whatever else. That's where I think we, as electrical engineers, get superpowers. Right?

Sergiy Nestorenko
Where we start to very easily create much more sophisticated boards, without, you know, this 3 month grueling process with 30 people reviewing it and and still letting errors go by.

Stephen Kraig
You know, from a a nuts and bolts perspective, every time I I'm we're saying, you know, we're taking this into account or we're considering this parameter, that's all something that seems to make the calculation explode into something much much larger. Right? Because a traditional autorouter might work in a sequential manner where it says connect a to b and then go to the next trace. Connect a to b and then compare. But if you're having to do all of this in parallel and consider how trace a and trace b talk to each other or don't talk to each other or whatever you're trying to accomplish, doesn't the problem become so large that it's incomputable?

Sergiy Nestorenko
So, I mean, in general, even the simpler version of this question. Right? Even, like, you know, bin packing or routing in a optimal way is an NP hard problem. Right? So it already starting to explode on you, you know, exponentially, and you don't have the hope for for an optimal solution.

Sergiy Nestorenko
So there's two answers to this. I mean, the first is simply that as humans, we don't find the optimal answer. Right? Like, we are not actually holding ourselves to, like, hey. This board is going to be the best possible set of atoms that could have created this schematic.

Sergiy Nestorenko
We're really just going for, does it work? And does it work to meet all of our specs? So the problem isn't to find the absolute optimum, typically. It's to find something sufficient for for your specs. But, really, that still explodes.

Sergiy Nestorenko
There's still a lot of considerations. There's still a lot of stuff to think about. That is what neural networks are great for. And they're not some sort of black magic that just, like, voodoo voodoo solve everything. That's not really what they are.

Sergiy Nestorenko
The way that I like to think of them is still a lot like caching or like memory. So what you can do is you can go and create millions or tens of millions or 100 of millions of boards. I'm thinking of the process here called reinforcement learning, which is how we approach this problem. And maybe the most famous example that people might lean to is what DeepMind did with the game of Go, where they beat the Go world champion in that game. The way that works is you have this neural network playing the game of Go over and over and over and over and over again.

Sergiy Nestorenko
And over at first, it's really dumb and making stupid decisions and could be beaten by a child, but eventually, it learns better and better and better strategies. And what it's really doing is it's not like recomputing the whole tree of all possibilities in Go. That is exponentially exploding and and not computable. What it's really doing is is it's kind of storing all of that previous experience in the memory of the neural network in such a way that you can interpolate previously unseen positions and make good decisions about them. So it's kind of like imagine, like, a big storage cache with interpolation capability.

Sergiy Nestorenko
So that's the way that we kind of solve this problem is by generating, you know, millions and millions and millions of boards in this reinforcement learning loop with all the physics simulations, which is too expensive for any one board, we can distill that into a neural network that kind of has that experience along for the ride when it creates the next one.

Stephen Kraig
So you've created just a really, really smart senior engineer that has all of this in their memory.

Sergiy Nestorenko
Right? I let's say we will create that really smart senior engineer with all that I remember. The the way that we break it down for now, just practically to make progress in this is, like, the most extreme version of this in a 100 years will be, like, a single neural network that takes a look at the schematic and, like, directly outputs copper. And that's where we might get into these, like, crazy shapes and these unpredictable things and blah blah blah. That is not how we do this today.

Sergiy Nestorenko
That would not work. No chance. What we do instead is is a much more practical approach. Right? So we take a lot of, you know, very classically written algorithms, like, okay, like, how do we detect collisions between 2 components?

Sergiy Nestorenko
How do we detect collisions between traces? You know, how do we find a path for a trace even in a classic way? And then we allow the neural network to make high level decisions. So it's almost like you have a senior engineer standing over the shoulder of the junior engineer saying, hey. You should put this component here and this component here, and then I'm gonna go have some coffee, figure out the rest.

Sergiy Nestorenko
And so the neural net is kind of that senior engineer giving that high level guidance. And then plenty of classical algorithms are still doing the low level stuff that, neural networks don't yet have the perfect precision for that would be required. But over time, like, this is the approach that Tesla took with autopilot. In the 1st days, it was just a very basic net looking at a picture and millions of lines of c plus plus driving it. And over time, the neural network became capable of eating away more and more and more of the c plus plus and made a better and better better system for driving cars.

Sergiy Nestorenko
And I we are very much in that progression now, and we deleted 500 lines of code just today with this kind of thing.

Parker Dillmann
That that deleting lines of code is there's a software developer at macro fab that that's his favorite job is when they make a commit, a GitHub commit, and, like, it just removes lines of code.

Sergiy Nestorenko
Yeah. Yeah. No. That that's honestly the best thing. Right?

Sergiy Nestorenko
Yeah.

Parker Dillmann
It's less maintenance.

Sergiy Nestorenko
Yeah. Like stealing this from Elon. Right? Like, the best line of code is no line of code. There's no unit test you need for it.

Sergiy Nestorenko
There's no regression test run. You don't have to debug it. You don't have to document it. You don't have to comment it. It takes zero time to run.

Sergiy Nestorenko
Like, if you can delete the line of code, still get your job done. Delete the line of code.

Parker Dillmann
So how does Quilter work today now? So if one of our listeners goes to quilter dotai, which is the website, this link will be in the show notes as well. What's their process? What do they are gonna be working through?

Sergiy Nestorenko
So the way we position this now is a lot like if you had a professional layup person that you were gonna email the board to. So roughly speaking, the way it works is like this. You do your schematic. That's still on you. You select your footprints.

Sergiy Nestorenko
We don't do any of that. You load that netlist and those footprints into the board file. Then it's a good idea to give us some constraints. So the first constraint that we must have is, like, what's what is the shape of your board? Right?

Sergiy Nestorenko
Is it a rectangle? What size is it instead of cutouts? Great. The second constraint you should probably give us is you should probably place your connectors or any buttons or any screens or anything that's, like, mechanically defined that AI has no business doing. You place that, and just leave everything else off the board.

Sergiy Nestorenko
Right? Like, don't worry about your stack up. Don't worry about your placement of the components. Don't worry about any of that. You basically drag and drop that file into Quilter.

Sergiy Nestorenko
So if you're working on Altium, you just drag and drop the Altium file, the PCB dock. If you're working in KiCad, then it's a KiCad PCB or whatever along with your schematic. Quilter will then ask you some questions about the constraints we can consider today. So for today, it's, like, live deployed. It's just power.

Sergiy Nestorenko
Right? So you will go through and define for these nets, I have 1 amps, 2 amps, or I really want, like, a specific trace width for some reason. Great. When you click run, Voltor will pass that off to a big cluster, and a bunch of GPUs, bunch of CPUs, and go spin and do stuff. And we will generate a whole bunch of candidates.

Sergiy Nestorenko
And so these candidates, you'll probably see 50, 60, 70 possible layout solutions. Some of them might not have made it to a 100%, and we'll kind of hide those. You can see them if you want to, but who cares about that? And then we'll float to the top the ones that we think are the best. So they got to a 100%.

Sergiy Nestorenko
And then what we also do is we simultaneously compile for as many manufacturers as we can. So obviously, like in some boards, you have to go down to these 2 mil, 3 mil, super fine pitch, like, specs that only a few manufacturers can handle. And some boards, like, okay, 10 mils for everything. Anybody can fab it. Hurrah.

Sergiy Nestorenko
And so we'll try all of the above, and whichever solution we got that is as flexible as possible will float to the top. And what you get is you just click a download button, and you're back in Altium or you're back in or you're back in KiCad. So at that point, you can open up the board. I I would personally encourage you to to run the DRC. If you see anything wonky, let us know.

Sergiy Nestorenko
But, hopefully, everything's okay, and you can proceed to submit to FAB.

Stephen Kraig
So what are some of the, challenges that come along with trying to do this?

Sergiy Nestorenko
Oh, man. I don't even know where to begin. There's a lot of challenges. On just the business side. Right?

Sergiy Nestorenko
You know, we've turned this into a venture backed company by raising money and convincing investors. Historically, it's been difficult to do. Right? Historically, investors have considered EDA as a small market, and especially when it's applied to PCBs and not chips. And so just getting sufficient funding for something that is this difficult And, you know, why is there such a proliferation of tools in the software space?

Sergiy Nestorenko
And, you know, why is there such a proliferation of tools in the software space that hasn't happened in electronics? And I think the answer is that in software, the people who use those tools can build them for themselves. Whereas electrical engineers are not typically programmers or not typically, like, amazing programmers. Right? So now our team has a whole bunch of folks who are amazing at software, but don't know much about electronics.

Sergiy Nestorenko
And then people who are amazing at electronics, but don't write a whole lot of software. And so, like, gluing that social world together is a challenge and is very important, but very exciting and very fun. And then technically speaking, there's a lot of challenges. Right? I mean, all of this stuff is is difficult.

Sergiy Nestorenko
You know, computational geometries is, deceptively difficult to do correctly. Writing the basic operations and then making them fast is okay, a straightforward engineering task. But, like, trying to come up with strategies and planning and figuring out how to train the neural nets in ways that they actually learn. I don't know how to say it. It's just difficult.

Sergiy Nestorenko
You tried things and they don't work a lot. And you try again and they don't work a lot. And you read papers and you try those things and those don't work either. And it's it's just a slog, but we're making progress on it for sure.

Parker Dillmann
How do the engineers or your community, like, how do they give you feedback? Is there, like, just a thought is there a like button? Is there a thumbs down? How how does that work?

Sergiy Nestorenko
So we we encourage as much feedback as possible. Right? Quilter is in open beta. Anybody's welcome to come use it. There's no restrictions.

Sergiy Nestorenko
You don't have to pay for it. Like, please go break it. Great. We have a chat on every single window. You can open it.

Sergiy Nestorenko
You'll see one of us, and you can tell us how much you like it or hate it immediately. We have little buttons that are specifically, like, please give us feedback. We have a community forum. We're hoping to start a Discord here pretty soon. We are as well I mean, like and we also exist on on Twitter and on LinkedIn and whatever else.

Sergiy Nestorenko
Like, please come give us constructive criticism left and right. The main thing we're doing right now is looking for people who also believe in this long term vision and wanna help us make it so. So we encourage in every way. It's funny. We we've thought about doing some sort of, like, ranking or thumbs up, thumbs down or something like that, like a hot or not for boards.

Sergiy Nestorenko
We haven't done it yet, but, it's a good idea.

Stephen Kraig
It's come up before. Swipe right for good layouts.

Sergiy Nestorenko
Yeah. There you go.

Parker Dillmann
Now how does that feedback into training the the neural net then? It's clear that's probably not like the user isn't directly doing that part of it. So how do y'all do that on your end?

Sergiy Nestorenko
I mean, the other consideration here is that, you know, people are generally sensitive about their designs, and you would hate to, like, have, you know, upload your design, have us train on it, and suddenly make decisions on somebody else's board based on your design. Like, that's not cool. So we don't do that. That's one of the nice things about reinforcement learning is is we actually don't learn from any human data whatsoever. We are just taking in, here is this game of playing PCB design.

Sergiy Nestorenko
Go play it, agent, and then we score you based on what matters, which is, is this board manufacturable? Do the physics checkout. Is it gonna work? Right? That's really what matters is, like, did you make this board manufacturable, and is this board going to work when you build it?

Sergiy Nestorenko
That's the score. Now there is an interesting debate to be had about, should we be trying to make these boards aesthetic? Right? I would argue that in most boards, that doesn't matter. Maybe if you're NVIDIA or Apple and your board is gonna be taken apart, people are gonna take pictures of it.

Sergiy Nestorenko
Like, the top and bottom layer, it's important to be aesthetically pleasing. But but in general, that shouldn't matter. But that's not something we've solved. Like, we're we don't have a beauty score that feeds into the system at this point, and and that's evident in the results.

Parker Dillmann
Before I was introduced to you, Sergei, through email, some people in a card Discord were using Quilter.

Sergiy Nestorenko
No way. I didn't know that.

Parker Dillmann
Yeah. And I saw the outputs, and I'm, like, that looks like a board from, like, the sixties in terms of, like, it's just organically just like a tree. And everyone was like, this looks really weird. And I'm like, well, this is, like, kinda like how boards used to look like because, you know, most people won't realize like, Steve and I are too young to have done boards that way. And, I've only had that experience because I've taken apart old stuff.

Parker Dillmann
And Steven probably because of amplifiers. Old amplifiers are that way.

Stephen Kraig
Well, I yeah. I've seen a good bit of it. And, I I hand etched my college project, so it had some organic issues. Yeah.

Sergiy Nestorenko
You you know, you you

Stephen Kraig
mentioned you mentioned Apple needing something aesthetic. I bet if you did do the curved trace thing or whatever with with Apple, there would be headlines saying, like, electronics reimagined or they've they've reinvented electronics going back to back to their roots or whatever. You know, earlier you mentioned that, Maxwell's equations. Are you actually computing Maxwell's equations for across the

Sergiy Nestorenko
board? To date, yes. And I consider it a mistake, unfortunately. So in in the long term, I think that's what we should be doing. Mhmm.

Sergiy Nestorenko
Right? In the long term like, the Maxwell equation is the source of truth. Right?

Stephen Kraig
So we should just

Sergiy Nestorenko
solve them properly, and let's not deal with any approximations and where those approximations might be faulty and all that stuff. That's the way we approached it. Like, our first solver that we plugged in, that we, you know, trained nets on at some point and experimented with was a full solver. Right? Like, it's a full, no approximations made, like, you know, f d FDTD type solver.

Sergiy Nestorenko
The problem is it's just too slow. Mhmm. That's the unfortunate thing. Like, when you have to take 5 to 6 minutes for a single query on even a small board, it's just not quite good enough. So we are now we can't use it as much as we want to in the solver.

Sergiy Nestorenko
And for that reason, we are putting a pause on that one and considering, like, alright. Well, why don't we go back to, the quasi static approximation? That's probably gonna be the next thing that we do is just alright. Like, basically, assume the time doesn't exist, and let's extract capacitance and mutual inductance, you know, and steady state currents, and and let's see how far we can get with that. And I I think that there's it's a valid approach.

Sergiy Nestorenko
I think we can start with that. We can train the agents to be pretty darn good with that and then fine tune with this more expensive full Maxwell equation solution later on. But yeah. But as a physicist and a purist, I wanted to go for the full Maxwell equation solution, and that was probably not the right decision at first.

Stephen Kraig
So let's say Quilter takes over the world of electrical engineering, and everyone is is on board with this. I noticed on your website, you had a a note on there that said that like, engineers, hardware engineers spend 50% of their time doing layout. And if they're not doing that, would how do you see the role of hardware engineering evolving?

Sergiy Nestorenko
I think, like, this question is is frequently asked out of the fear of, like, oh, no. What's gonna happen to my job? Like, what am I gonna do? Right? And I think, like, this question is gonna be really relevant in in the age of AI and in the age of automation.

Sergiy Nestorenko
And, historically, it's always turned out favorable. I mean, there's there's short term disruption. Right? Like, when, you know, you come in and your specific job is automated away, for for a specific person, that might be difficult. I don't think Quilter is gonna, like, suddenly wake up tomorrow and have a GPT 4 moment and everything is doomed.

Sergiy Nestorenko
Like, I I

Stephen Kraig
you're fine. Me we can't party now 50% of our time?

Sergiy Nestorenko
No. No. If you only design boards that are, like, sub 2,000 pins and sub 500 megahertz, then sure. But in general, no. But, you know, I think that there's today already such a shortage of electrical engineering talent.

Sergiy Nestorenko
Right? Like, most a lot of people are graduating college and being like, screw this. I'm gonna go learn software because it pays better and it's easier. And so the hardware companies are stuck. Like, you've got, you know, an aging population who's getting ready to retire in these fields.

Sergiy Nestorenko
Not that many people coming in. I just saw this morning, I think it was the CEO of Intel a couple years ago posted a chart of, like, you know, in 1980, you know, like, everybody was graduating with an EE degree, like, in this space, and, like, 10% of people were a CS degree. Now it's exactly the opposite. Right? Like, there's, like, 10 times as many CS people as there are EE people.

Sergiy Nestorenko
So I think there's no shortage of work to go around. Right? Let's get all those people contributing to schematic side, getting boards done faster. You know, how many times at SpaceX did we have, you know, like, we were testing some chip and all we could do is, like, string together some breadboards and some breakout kits and some horrible stuff. And then, like, when you carry it from one lab to the next, some wire comes loose and you spend 2 hours debugging, like, why that happened and before you can do your test.

Sergiy Nestorenko
Like like, no. Like, like, now that it's easy, let's just make a schematic real quick and make throw together a board and then have it be reliable. You know, I think that that's what's gonna happen. Like, I think that those folks who are currently PCB layout experts can move into a schematic or go into firmware or go into test engineering or go into environments or any of the other fields that are critically important to designing PCBs, and we need them. There's a shortage.

Sergiy Nestorenko
Like, this is this is not gonna be like a lose your job, nothing for you to do. There's so many other things to do.

Parker Dillmann
Yeah. On top of that, it's the just the CHIPS Act alone in the United States is 600,000 engineering jobs that we need in the next 4 years. And we've talked about this previously on the podcast. It's on our current rates, it would take us 20 years to fill all those positions, and we have to do it in 4.

Sergiy Nestorenko
Wow.

Parker Dillmann
So it's like

Sergiy Nestorenko
I didn't know that.

Parker Dillmann
Yeah. It's it's even worse than you think, Sergei.

Stephen Kraig
I no. Thank you for telling me. I did not know that. That's remarkable. Yeah.

Stephen Kraig
We've been saying right now is a fantastic time to go get a double e degree because there's jobs waiting for you right at the end.

Parker Dillmann
Yeah. Yeah. And I I think that's why I I view AI as something that's super helpful and just helps you automate parts of your job that end up repeating. Like, like what you're saying is is routing a board. A computer should be really good at that.

Parker Dillmann
Right?

Sergiy Nestorenko
Absolutely. And, hey, I I know people who think of layout as therapeutic, who do layout as a as a way to kinda stay calm and and, you know, process their day. And you can always do that. Yeah. That's fine.

Sergiy Nestorenko
That's fine. But, but, yeah, in cases where, like, you need results, you need the board fast. I mean, at SpaceX, right, like, we valued one day if you could cut a day at a critical path by spending a $100,000, that was an easily smart decision. Like, obviously, go do that. Because one day of critical path is is much more expensive than that.

Sergiy Nestorenko
Right? And so if we're talking about cutting weeks or months out of critical path, it's it's a no brainer. You know, in those applications where engineers could be doing something better and can get the product out that much faster, like, of course, let's do this.

Parker Dillmann
You know, why vision here is because you you talked about how it culture will output multiple designs is I'm viewing this as you take this multiple designs and you build them, all of them. And then you validate them in real life with all your parameters, put them in all your test equipment and that kind of stuff. And that is super powerful, and that's something you just can't do right now. Right? Because that's that's the whole thing with software.

Parker Dillmann
We're talking about the compile button, right? Where it's so quickly to validate and change your design if it doesn't work. But in hardware space, it's minimal a couple of days just to get a piece of hardware. And usually it's like a week to 2 weeks to get a piece of hardware on your desk. So if you can go, hey, I wanted to I want to try 10 iterations right now.

Parker Dillmann
You can't do that.

Sergiy Nestorenko
I wanna get to the point where, like, we know this is the best candidate and it's gonna work. But until then, yeah, that's valid. That's a cool idea. I haven't thought about it that way.

Stephen Kraig
It's funny because I was just about to bring up the exact same thing, and this is something I've never had to do before. Just it just doesn't really make sense. But if you had 3 different layouts, let's say, and you wanted to build 30 boards, 1010 and 10 of each, they all have the same parts. It wouldn't surprise me if you could work a deal with your contract manufacturer where they leave all the parts on their machines and they just load this layout, do it, load the next layout, do it. You know, they do have to do a change over on the program and things like that, but you may save actually a little bit of money, and you could try 3 different layouts.

Sergiy Nestorenko
You you could do this one board. Right? Like, you could just make it a panel where, like, the panel has one of each shape. True. And then it's just 1 PCB placement pass, and then it's the same parts.

Sergiy Nestorenko
And, yeah, I mean, this is be a cool way to experiment with, like like, how small can we make a board? Right? Like, let's try 10 different sizes from fairly big to, like, as dense as we possibly make it. And then, like, what's the smallest one that actually works and is reliable and high yield and bring up? Like, okay.

Sergiy Nestorenko
That's the one we're going with. Cool.

Stephen Kraig
As long as you have

Sergiy Nestorenko
the money

Stephen Kraig
to do it.

Parker Dillmann
So, for those interested in trying quilter, Sergei, where can people go and get started working on it?

Sergiy Nestorenko
Yeah. Easiest place is Quilter dotai in your browser. If you search for quilterpcvai something something in Google, you'll definitely find us. You know, if you search for my name on on LinkedIn or Twitter, you'll find me there as well, and the links lead back to it.

Parker Dillmann
Again, thank you so much for coming on podcast and talking about Quilter.

Sergiy Nestorenko
Oh, thank you, guys. Steven Parker, it was a pleasure. Yeah. Thank you. Thanks so much.

Parker Dillmann
Thank you for listening to circuit break. We're your hosts, Parker Doleman.

Stephen Kraig
And Steven Craig. Later, everyone. Take it easy.

Parker Dillmann
Thank you. Yes. You are listener for downloading our podcast. Tell your friends and coworkers about Circuit Break, the podcast from MacroFab. If you have a cool idea, project, or topic you want us to discuss, let Steven and I and the community know.

Parker Dillmann
Our community where you can find personal projects, discussions about the podcast, engineering topics, and news is located atform.macrofabdot

Sergiy Nestorenko
com.

Related Podcasts

CB FI 427

Food Device Contest Wrap up

We celebrate the innovation and creativity of Food Device Design Derby entries like the JavAqua, Pizza Pouch, and the winner, BarBuddy.

CB FI 425

MacroFab Platform Updates from Kyle McLeod and Nicholas Lundgaard

The episode provides insights into MacroFab's efforts to make PCB manufacturing more accessible and efficient for their customers.

The pcb plague

The PCB Plague

Ever have PCBs that solder just will not wet and solder to? You probably thought it was improper soldering technique but that was probably not it!

MEP FI 406

Dr. Duncan Haldane from JITX on Automating Circuit Design

Dr. Haldane on his background, the problems JITX is trying to resolve, & new auto-router plans. What's the deal with the "hyper-aggressive pogo-stick"?

Breadboarding for success

Breadboarding for Success

This week we are talking about Breadboards. Is breadboarding a circuit or design still applicable in today's SMT component dominated world?

Datasheet lore

Datasheet Lore

What lore have you discovered in component datasheets? On this episode, Parker talks about how he picks electrical components and risk management.