I was working on an ADC design recently and realized something uncomfortable: a lot of what I “knew” about analog noise came from blog posts and forum advice, not from anything I’d measured myself. I’d repeat things — keep the switcher away from the analog rail, don’t run digital traces under the ADC — without having seen what those rules actually buy you when you break them.
So I’m building a board to find out.
The approach#
Rev 1 is what I’m calling a zero-thought design. I didn’t curate a list of textbook mistakes. I just laid the board out without thinking about any of the things analog designers tell you to think about — SMPS sitting next to the ADC, no real plane strategy, decoupling chosen the way I’d pick it if I didn’t know better. Whatever falls out of that is the starting point.
From there:
- Measure the noise on rev 1 — RMS noise floor and an FFT of the output.
- Attribute each noise source to a specific design choice, using a mix of bench measurement (changing one thing at a time) and datasheet math — PSRR, reference noise specs, that kind of thing.
- In rev 1, fix one thing. Measure again. See if the prediction held.
- Repeat.
One change at a time is the point. I want each fix to be a small experiment, not a redesign. If the prediction was wrong, I want to know — and that’s only possible if I haven’t changed five things at once.
What I expect to find#
I don’t know yet. Some rules will hold up. Some will matter less than I expect. Some will matter in ways I didn’t see coming. If I knew the answers I wouldn’t be doing the project.
There’s also a real question about whether you can cleanly isolate causes when you’ve broken everything at once. I’ll come back to that in the next post.