Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use High-Speed PCB Design to Beat Noise
Transcript
- Lucas: So you spend months on a prototype, get the boards back, power it up — and the Wi-Fi barely works. Range is half what you expected. You swap antennas, no change. You call the module vendor, they blame your layout. Luna: Classic. And nine times out of ten, it's not the module. It's a signal integrity issue on the PCB. Lucas: Right. And this is the episode where we talk about high-speed PCB design — because once your microcontroller runs above fifty megahertz, or you have a radio that transmits at hundreds of megahertz, the board stops being a set of wires. It becomes a transmission line problem. Luna: And most hardware founders I meet treat the PCB layout like it's just connecting dots. They hand it to a junior engineer or an overseas layout house with no guidance. Lucas: Yeah. And the result is what I saw at a medical device startup last year. They were building a continuous glucose monitor — small board, Bluetooth Low Energy radio, plus a sensor analog front-end. The first spin passed functional test in the lab but failed radiated emissions by a huge margin. They spent six weeks and twenty thousand dollars on re-spins. Luna: Ouch. What was the root cause? Lucas: Three millimeter stub on the antenna feed line. The layout guy routed the trace from the module pin to the antenna connector, but he put a tiny test point pad inline — just three millimeters off the main path. At 2.4 gigahertz, that stub acts like a quarter-wave radiator. The board was essentially broadcasting noise at exactly the wrong frequencies. Luna: I've seen almost the exact same thing at a drone startup I worked with. Our compass kept giving garbage readings. Turned out the I2C lines from the magnetometer ran parallel to a high-current motor PWM trace for about twelve millimeters. No ground plane between them. The coupling was brutal. Lucas: That's the second most common mistake — not having a continuous ground plane underneath critical traces. A lot of people think ground planes are just for grounding. They're really for controlling impedance and providing a return path. If you split the ground plane with a gap, the return current has to go around it, which creates a loop antenna. Luna: And loops radiate. So you fail emissions. Or you get crosstalk that corrupts your sensor data. Lucas: Exactly. The third big one is ignoring stack-up symmetry. In a four-layer board, if you put signal layers on top and bottom and power and ground in the middle, you need the prepreg thicknesses to be balanced. Otherwise the board warps during reflow, and you get micro-cracks in vias. But more subtle: if the stack-up isn't symmetric, the impedance of traces changes from one layer to the next. Your fifty-ohm trace on layer one might be fifty-two ohms on layer four. Luna: And then your impedance mismatch causes reflections. Which messes up signal integrity again. Lucas: Right. So the core idea is: treat every trace that carries a clock or a high-frequency signal as a transmission line. That means you control its characteristic impedance — usually fifty ohms for single-ended, ninety or a hundred for differential pairs. And you terminate it properly. Otherwise, the signal bounces at the receiver, creates ringing, and that ringing couples into other traces. Luna: Can you give a concrete example of what proper termination looks like for a hardware startup? Lucas: Sure. Simplest case: a point to point SPI clock line running at twenty megahertz. If the trace is longer than about two inches, you need a series resistor at the source — typically twenty-two to thirty-three ohms — to match the driver's output impedance to the trace impedance. That damps the reflections. I've seen startups skip that resistor to save two cents. Then they wonder why their display glitches. Luna: Two cents. And it costs them a week of debugging. Lucas: Yeah. And the tools to catch these issues before you fab are cheap or free. KiCad has a field solver plugin. There are free online impedance calculators from JLCPCB and PCBWay. But I think the real problem is mindset: founders think of PCB layout as a software task — you just place components and autoroute. It's not. It's a physics problem. Luna: So what's the one piece of advice you'd give to a hardware startup on their first board? Lucas: Before you send the Gerbers to fab, run a two-dimensional field solver on every critical net. That means your clock lines, your differential pairs, your antenna feed, your USB data lines. The field solver will tell you the actual impedance and show you where the return path is broken. It takes maybe an hour. It will save you months. Luna: And if you don't have someone on the team who knows how to do that? Lucas: Hire a consultant for a day. Seriously. A thousand dollars for a layout review is cheap compared to three board spins at ten grand each. And the consultant will also catch stuff like via stub length and cross-talk between aggressor and victim traces. Luna: Let's talk about vias for a second. Because I've seen startups use via-in-pad on fine-pitch BGA packages to save space, and then they have solder voiding issues. Lucas: Yeah, via-in-pad can be a trap. If you fill and planarize the via properly, it's fine. But if you just tent the solder mask over it, air gets trapped and expands during reflow, causing voids. And voids increase resistance and reduce reliability. The better approach for low-volume prototypes is to move the via outside the pad and use a short trace to connect. Luna: Or use microvias if your fab can do them. But at a hobbyist price point, microvias aren't always an option. Lucas: Right. So back to the medical device startup story. After fixing that antenna stub, they also changed their stack-up. They had originally used a two-layer board because it was cheaper. But at 2.4 gigahertz, two-layer boards are almost impossible to get right — the return path is too far away, so the trace impedance is uncontrolled. They switched to a four-layer board with a ground plane on layer two. Cost went up maybe fifteen percent. Emissions passed first try. Luna: Fifteen percent more on board cost, versus six weeks of schedule delay. No contest. Lucas: And the thing is, the delay kills momentum. If you're a startup raising a seed round, a six-week delay can mean missing a demo day, missing a customer eval, missing a grant deadline. The board cost is irrelevant. Luna: So let's run through a quick checklist for anyone listening who's about to do their first high-speed layout. What are the top three things to verify? Lucas: One: do you have a continuous ground plane directly under every high-speed signal layer? Two: is every high-speed trace length-matched if it's part of a differential pair or a parallel bus? Three: have you simulated the impedance of your antenna feed line and verified it's fifty ohms at the frequency of interest? Luna: And a bonus one: do you have proper decoupling capacitors near every power pin, with the smallest capacitance value closest to the pin? Lucas: Good catch. Decoupling is its own topic, but it's related. If the power supply has high-frequency ripple because you skimped on decoupling, that noise couples into your signal traces. Luna: You know, these conversations that we have on the show — they're the kind of stuff that I wish I'd heard three years ago when I started in hardware. It would have saved me a lot of burnt weekends. Lucas: Same. And honestly, that's why we keep doing this. A couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah. It's a small thing that makes a real difference. Lets us stay independent and dig into topics like this without ads. Lucas: Now back to vias — another common pitfall is the via stub. On a multi-layer board, if you route a trace on layer one and use a via to go to layer three, the unused portion of the via barrel from layer three to the bottom acts as a stub. At high frequencies, that stub resonates and creates a notch filter. You can lose three to five dB of signal at exactly the frequency you care about. Luna: How do you fix that? Back-drilling? Lucas: Back-drilling is the proper solution, but it adds cost. For low-volume prototypes, you can avoid the problem by routing critical signals on the same layer or using blind vias if your fab supports them. Or you accept the loss and design around it, but at least be aware. Luna: I've also seen startups try to save money by using a single ground plane for both analog and digital sections. That's a disaster for mixed-signal boards. Lucas: Yeah. You need to partition the ground plane — physically separate analog and digital grounds, and connect them at a single point, usually at the ADC or the power supply. Otherwise, digital switching noise injects straight into your analog front-end. Luna: So the takeaway for a founder listening: don't treat PCB layout as an afterthought. It's a core engineering discipline, and mistakes there are expensive. Lucas: And the good news is, the tools and knowledge are more accessible than ever. There are open-source field solvers, free impedance calculators, and a community of engineers sharing their failures. You just have to invest the time to learn. Luna: Any final thought? Lucas: Yeah. The next time you're about to click 'order' on a prototype, stop. Run that field solver. Look at your ground return path. Ask yourself: is that antenna feed line really fifty ohms? It might save you a very expensive lesson.