Integrated Circuit design: Comparing the results given by corner analysis and Monte Carlo analysis

Question

In IC design, corner analysis is performed to study the operation of the ICs under extreme conditions and Monte Carlo analysis is performed to ensure the mismatch is within the tolerance.

However, can the results given by the corner analysis and the Monte Carlo analysis be different? For exapmle, in a TT corner @27C, the output voltage is, say, 1.8V. But the mean output voltage given by the Monte Carlo analysis @ 27C is only 1.6V even if the corner analysis and the Monte Carlo analysis are based on the same set of design parameters (such as L and W).

Also if this happens, what would be the solution? Would different sets of parameters be used for the corner analysis and the Monte Carlo analysis?

Thanks

Answer 1

I'm not an expert in variability analysis, but Monte Carlo and corner are looking at two different things.

Corner analysis is looking at worst case (from a digital perspective mostly): All your N/P devices are at the 'slowest' point possible, or are at the 'fastest' point possible. I'm mostly familiar with corner analysis for digital designs to verify that, in the worst cases, all timing is still met (because if it is met under those situations, it will also be met on any intermediate situation).

Monte-Carlo looks at variability of devices within the IC (and depending on how it is run, also looks at resistor/capacitor matching, wheras corner-study only looks at the active switches). Things like how much offset you can expect from OTAs, how much differential-to-common mode conversion, how well matched are your gains of different amplifiers, etc etc.

So yes, they can give you different numbers (and probably should, really), because corner analysis looks at one specific, wafer-level device 'variability point'. Monte-carlo studies inter-wafer mismatch and so on.

Answer 2

The corner analysis results will depend on what your device_characterization people embed into the corners.

If interactions are ignored, then the corners will be beyond worst_possible.

So get to know the foundary people who characterize the process output, usually on special test wafers, and ask about the tradeoffs they embedded into the corner models.