AD8310/AD8307 oscillation/ripple with RC discharge shape


I am trying to use AD8307/AD8310 (I tried both on different boards) to detect the shape of 50-400us pulses (DC pulse). I am using the chips in DC connected configuration, signal placed at the bias voltage advertised in the datasheet (3.1V for 5V supply), large 10uF capacitor on offset pin and connected to a 12 bit 2MSPS ADC, buffer at input and output before ADC. Everything is working fine as expected except random ripple/oscillation on both chips AD8307 and AD8310.

The oscilations have an unusual/unexpected shape and resemble a RC discharge curve with rapid buildup and in my case discharge with tau around 1 us in my case. By unusual I mean, that if the whole chain of the amplifier cells in the chip goes to oscillation then I would expect either something like "log sin(t)" shape or at least sin shape if none of the cells goes into saturation/limiting. Which leads me to the conclusion that the chain of amplifiers is not oscilating, but rather something else in the chip and there are two possibilities - either the output interface or the offset, where both have at least some capacitance, which could cause this behaviour. By the way, the feature is a bit chaotic, frequency depends on level of signal (low signal clearly shows more oscillation, but I am getting also bursts of oscilations at input level in 10's of mV - burst = they appear and go away in matter of say 10-15 us, whereas at low level signal I can have sustained oscillations, the frequency is typically somewhere between 300-500 kHz, but changes). 

I was reading through this site and it seems that I am not the only one observing this behaviour and I am linking just a couple for reference:

For me the problem is that the oscillations destroy the response at low levels and basically make the design/use of these parts not feasible. So my main question is - do you know what might be the likely source of the problem? I.e. source/reason for the RC ripple.


Edit 2020.03.25 - JD
[edited by: jdobler at 11:53 PM (GMT 0) on 25 Mar 2020]