Does MC mismatch simulation reflect best or arbitrary layout ?

Mismatch simulation actually draws no conclusion about the quality of the layout - what you are simulating is not the systematic mismatch due to poor alignment, but the residual random variation of each local device. 

So put another way, the models are not trying to model the effects caused by good/bad layout. 

There may then be additional mismatch caused by bad layout, but that's not what you're simulating. If that was what was going to be simulated, the foundries would have to characterise this somehow and then there would need to be a tool which tried to assess correlation based on layout position - there are no models and no tool to do that though. There are some LDE (layout dependent effects) analysis tools which can look at the layout environment and use that to capture the change in behaviour of devices due to their environment, but that's not quite what you're asking here.

That said, running Monte Carlo with mismatch does tend (particularly if you use mismatch contribution) to give you a feel for the variability of which devices leads to the biggest variation of your output measurements which could help you assess which devices require the most care about placement.

Andrew.

Mismatch simulation actually draws no conclusion about the quality of the layout - what you are simulating is not the systematic mismatch due to poor alignment, but the residual random variation of each local device. 

So put another way, the models are not trying to model the effects caused by good/bad layout. 

There may then be additional mismatch caused by bad layout, but that's not what you're simulating. If that was what was going to be simulated, the foundries would have to characterise this somehow and then there would need to be a tool which tried to assess correlation based on layout position - there are no models and no tool to do that though. There are some LDE (layout dependent effects) analysis tools which can look at the layout environment and use that to capture the change in behaviour of devices due to their environment, but that's not quite what you're asking here.

That said, running Monte Carlo with mismatch does tend (particularly if you use mismatch contribution) to give you a feel for the variability of which devices leads to the biggest variation of your output measurements which could help you assess which devices require the most care about placement.

Andrew.

Hmmm, I understand that that the simulator cannot draw any conclusion on the quality of the layout.

Basically you are saying that some effects (LDE:  poly/doping/other gradients along the wafer) are not included in the mismatch simulation - and a care must be taken in layout to compensate these (unmodeled) effects ?

This actually tells me, that it is always worth to do best common centroid for matching devices, because additional effect will come on top of the mismatch parameter variattion.

Just after posting this question I found the relevant chapter in the DRM that states: "the mismatch model includes the analysis of identical devices IN CLOSE PROXIMITY".

That tells me that the mismatch (and thus the measured results) can be worse if I do not take care about proper layout (centroid), dummies, etc. 

Did this summary match what you wrote ?

Effectively, yes. You can model some of the LDE effects by extracting relevant parameters (e.g. LOD, stress parameters such as SA, SB etc, well proximity effects) but this is not really anything to do with the mismatch simulation. The foundry document you found is effectively saying the same thing - that the model is just capturing the remaining variation (the random part) after good layout practice has been taken into account (that variation would be systematic).

So yes, what you're saying is effective the same as what I'm saying.

Andrew.

Perfect !

Thanks once again !