Practical PCB RF design
Wow, one of the few times the YouTube algorithm has recommended an extremely useful video.
In a few minutes I learned more about practical PCB RF design than I ever knew. Very easy to understand: https://www.youtube.com/watch?v=xhuHAhIKWoM
Comments
He would have done even better to pay attention to making the 50 ohm traces, indicated by the presence of the 50 ohm SMA, be 50 ohms. For FR4 that probably indicates a width/height of a bit under 2.
QUCS has a tool for this.
Including the coupling lines Z and L in the predistorted design would help too.
Agree.
One thing for me that I didn't fully appreciate is how the stackup can affect loop area for high frequencies more than you might think. Especially since the inter-plane distances can be so much less in the prepreg (e.g. top to L2) compared to the core (e.g. Top to L4 for a 6-layer board). Where a trace from Top to L4 is jumping across the core distance between L3/L4, creating a much larger loop.
Not sure how you feel about this, however I watched an interesting video on it.
https://www.youtube.com/watch?v=vvURxAfMPT0
Yeah, it's funny you mention AI and PCB design.
Just yesterday (not kidding) I was in KiCAD and had to put together a very large BGA footprint for a part not in the KiCAD standard footprint library. It was moderately annoying because out of the square ball grid array many of the ball positions are missing on this particular package. I had to consult the data sheet carefully and make sure I was deleting the right ones.
Part way through that tedious and error-prone process I thought, hey, I wonder if AI can do this for me? So I entered a prompt in X-Grok (the only AI I've ever used) and asked for a Python script to do just that: Create a KiCAD footprint from the part data sheet. It did a pretty good job. But I kept having to refine the prompt because it got confused about the package characteristics between two possible package types. Even though I told it exactly which one of the two to use.
In the end, after everything was right, I got to the bottom of the code it produced and it said: "Now you'll have to go to the data sheet and manually extract a list of the array positions to exclude from the result, and enter them here in the code". The very thing I wanted it to do! What a waste of time!!
I went back to KiCAD and finished making my footprint in less time than I spent messing with AI 🙄
I only use a paid for version of ChatGPT - and use it for some daily tasks like writing SQL queries, some dirty python, HTML etc, and reviewing texts, all the stuff I don't have the bandwidth for normally.
I get what you are saying about having to refine your prompts constantly, however in the above cases, it's a lot less time for a positive outcome rather than a) paying someone else, or b) a steep learning curve for something I only have limited knowledge on.
I might see how ChatGPT can interpret and produce a board layout in KiCAD - something I haven't used in anger before, however, might be time to start learning. I haven't signed up for Flux.ai yet either, so will give it a try too - from a complete newbie perspective.
Hi Glenn,
I thought he was initially making a deliberately bad layout, to emphasise the design aspects.
But maybe I misunderstood the part you were referring to.
Regards,
Martin
It was a filter built in a 50 ohm context, as demonstrated by 50 ohm SMA connectors. The lines from the connectors to the filter are not 50 ohm but something a good deal higher Z.
Relatedly the lines between even his physically isolated resonators are separated significantly by higher Z microstrip. This isn't necessarily wrong but needs to be included in a design and implementation.
To make the best filter not only do the input lines need to be appropriate to a matched [50] ohm filter design but the effects of the coupling lines should be absorbed into the design in order to get the expected results.
While he dealt with some of the grounding issues he did not address the line impedance question.
In an exhaustive design, not only the via/ground paths must be controlled and included but also the input and inter-resonator lines. Also more accurate models of the lumped elements need to be used, particularly for inductors, addressing the Q is very important. Usually C's have enough higher Q that L's become the bottleneck. He may have done this, he doesn't show his design model but no mention was made of line Z which not a minor factor.
With attention to these kinds of details, extremely good tracking between modeling and measurement is possible.