Electronic – Two same ADCs produce different result for same signal

Form what I read: "The digital ground pins (GND), the analog ground pins (AVSS), and the GND_CORE pins on the LAN8720 QFN are all connected internally to the exposed die paddle ground. "

So use a separate analog ground for your ADC as you mentioned above. Maybe it will work if both share the same ground plane, but it's easier to omit a ferrite if you don't need it than add one in if you do.

It's not exactly clear what your setup is here, but I've had a couple of applications that have done things like this. I can give you an idea of what the methods used were. Both of these solutions were in actual use, but had different requirements and methods.

1. The first method was to use a voltage controlled oscillator that drove my ADC clock for synchronized audio sampling. The idea is to have the oscillator run a counter and use some software to determine if the number of counts between PPS signals was off of expected. Then a control loop was implemented in firmware to adjust the voltage input of the oscillator (slowly to prevent artifacts). This created very good synchronization, but it does require specific hardware support, and the effects of the adjustment rate on the signal have to be well understood.

2. The second method was pure software based. In this method, the ADC clock was free running. This ran a free-running counter as well. When a PPS signal occured, the current counter value was stored as a data point inline in the ADC samples. Then the data from all the ADCs were combined by resampling them in software post processing to align the PPS sample timestamps. Again, this requires pretty decent knowledge of the signal so you an know which signal processing artifacts are acceptable and which aren't.

A way of looking at it is the first method is a control systems solution while the second is a digital signal processing solution. Both are valid, but both have pros and cons and the solution has to match the application needs.

For example, the first method requires more specialized hardware but much less processing power, while the second one can be implemented on more generalized processors. In my case, the first was an embedded unit that needed to run with fairly low power. The second was a data collection unit, so we could process the data later on a PC.