Electronics and PCB design concepts, explained in animation daily. Builder of @GetAlignedApp. LDS. KG7VYL
Washington, USA
Joined March 2022
- Tweets2.7K
- Following302
- Followers1K
- Likes16.3K
A 100 nF next to a 10 µF is classic decoupling, but the pair hides a peak. Between their two dips, the 10 µF already acts like an inductor while the 100 nF still acts like a capacitor, so they resonate.
In this model: 154 mΩ at 11 MHz, 5× higher than 10 µF + 1 µF.
When a high speed signal drops from the top layer to layer 3 through a normal via, the barrel doesn't stop at layer 3. The leftover bit below is a stub, and it's a dead end: part of every fast edge runs down it, bounces off the bottom and comes back as an echo.
That 4.7 µF ceramic cap on your 12 V rail probably isn't 4.7 µF. Ceramic caps lose capacitance under DC voltage, and the same part in a bigger package holds on to a lot more of it. At 3.3 V the effect is much smaller for parts rated well above that, like these 16 V caps.
If your board's clock drifts ahead, check the crystal's load caps.
A 32.768 kHz crystal whose datasheet says CL = 12.5 pF doesn't want 12.5 pF caps: they sit in series, so it sees half that plus stray. With typical numbers it runs 40 ppm fast, gaining 1min 44s per month.
I’ll say it. Patch panels are pointless and just decrease signal integrity.
It’s easy enough to reconfigure the VLAN for any output.
If you've put an ESP32 on a board, where it sits matters more than you'd think.
Espressif tested one of their modules two ways: antenna past the board edge, and mid-board over copper. On channel 11, the mid-board one put out 6.9 dB less, about a fifth of the power.
Ever had a ground pin that just won't solder?
The plane is pulling the heat away. Same iron on two pads tied to GND: the solid connection took 4.6 s to melt the solder. With a thermal relief (4 thin spokes), 0.25 s.
Your iron isn't weak. The plane is winning.
Here’s an actual simulation that describes this effect.
nitter.cf/michaelskywal/status/2…
Where you put a buck converter's input cap matters. Each time the chip switches, current jumps around a loop. A longer loop rings harder due to the increased inductance. Keeping the input capacitor close reduces the ringing.
An open via in an SMD pad is a tiny drain. A 0402 pad gets about 0.019 mm³ of solder from a 0.1 mm stencil, but a 0.25 mm via hole through a 1.6 mm board has room for 0.079 mm³, over 4× more. When the paste melts, the copper barrel wets and pulls the solder down.
A ground strip beside a fast trace only shields if it's stitched. With a 100 ps edge and a quiet trace 1 mm away, no strip lets 5.8% of the signal couple over. A strip with one via rings and makes it worse: 6.4%. Add a via every 5 mm and it drops to 2.8%.
Electrostatic discharges can easily damage sensitive semiconductors.
An 8 kV zap ramps to 30 A in under a nanosecond. Put the TVS on a 12 mm stub with a long run to ground, and that ~10 nH adds ~390 V arrives that at the chip.
A Diode right at the connector: ~47 V.
Your 1 oz copper trace is about 35 µm deep. At 1 GHz, the current lives in a thin layer about 2 µm deep.
This is called the skin effect. As frequency rises, current crowds toward the surface. By 1 GHz, only ~9% of the trace's copper is doing any work.
Your 100 nF decoupling cap stops being a capacitor at higher frequencies.
It filters best around 16 MHz. Above that, the inductance of its pads and vias takes over and it acts like an inductor. By 1 GHz it's about 200× worse than at its best.
What one button press looks like on an oscilloscope (simulated). Your code saw 6 presses.
Metal contacts bounce for about a millisecond before they settle, and the pin sees every bounce. A 100 nF cap can't recharge between bounces, so only one press gets through.
Your I2C pull-up sets your bus speed. Chips pull the line low fast, but only the resistor pulls it back up. Every pin and trace on the bus is a capacitor it has to charge.
In a differential pair, every bend puts one trace on the inside, which ends up shorter.
Two turns at 20 mil pitch ends up being about 63 mil shorter and ~10 ps early. Now the halves don't cancel, and that leaks out as noise and EMI.
A small serpentine adds the length back