Greetings

I am trying to simulate in Cadence a class-AB pseudo-differential (with a harmonic mean splitter) log-domain biquad. I am using AMS 0.35u CMOS process and the kit is c35b4b. All transistors are operating in the weak inversion regime.

I am performing a transient analysis to assess the filter's linearity at the output for a given input signal strength.

For very high input signals, greater than 30nA while my dc biasing is 1nA (where the internal non-linearity of the filter comes into play) and my settings set to conservative on gear2only, i get a very weird signal at the output - see graph1 attached (The two upper waveforms are the signals directly from the splitter that are fed into the two log-domain biquads and the other waveform is the output),...when i switch my integration method to Euler, i get the corrent result -see graph2 attached (but with lower point-to-point accuracy). If i use Euler with relaxed relative tolerance (reltol = 100e-5) i can increase the input strength even more (up to 50nA relative to 1nA DC bias).

None of the other integration methods give the expected results no matter what relative tolerance you set.

My problem is i dont know what to trust...is it because the system is inherently internally non-linear and the integration methods are 'inadequate' to give a solid solution for strict tolerances, am i pushing the models too hard or the filter just does not work?


Thanks a lot!