On the timeseries for Manual avg (grey) one can spot some drift upwards.
Yeah, makes me wonder what the auto-zeroing inside the ADC is actually doing. Also - the constant offset between the two traces?! Maybe still beneficial to have mux in the front end to do occasional zero measurement.
Any idea where the peaks ~11Hz & 22Hz originate from?
Nope. I suspect that in a run of 32768 samples, there will be a sample that is a massive deviation that drags it down/up. Note: I used the ADC test mode (where it outputs a constant value) to check that it isn't a glitch on the flexSPI readback.
How do you extract rms voltage noise frequency density from timeseries?
I have an NSD estimator algorithm that I have validated previously (to ensure the amplitude is "correct"). Uses multi-taper spectral estimation with discrete prolate spheroidal sequences (yes, a mouthful). It provides slightly better spectrums for the same amount of input data when compared to the traditional Welch method, not really enough of an improvement to justify it, it was just an intellectual curiosity for me.
How does your test setup look like: metal case, SPI isolator, battery powered PSU?
No metal case (tried it in my shielded steel experiment box, no change), no SPI isolator (only level translator), USB to laptop (switching laptop between AC and battery didn't make a difference). The analog supplies are powered from 9V battery, digital supplies from USB connection.
The data look like a result of a split-ground setup
Single ground-plane. As this is a BGA package, every ground pin goes through via to ground plane whereas I note, with interest, that the eval board uses solid copper planes directly under the BGA. I may try that, despite being against BGA-layout convention (would also mean double copper thickness compared to plane on internal layer). BGA fanout attached for your intellectual curiosity.