How many Monte Carlo runs should be done and what's the minimum success percentage?

I designed an analog circuit and I am running Monte Carlo analysis.
1. How many Monte Carlo runs are assumed to be sufficient?
2. What of the minimum percentage of successful runs (runs with good results) to consider the circuit is OK?

Monte Carlo runs

In general I would run 100. I usually calculate cp and cpk using the mean, limits, and standard deviation. If cp and cpk are > than 2 the design should be OK.

Re: Monte Carlo runs

Thank you Frankliner. The design is a sample and hold circuit.
If out of 1000 runs, I have, let's say, 50 runs where the output is not stable (oscillates) and the remaining runs result in good Harmonic distortion figure.
Is that considered acceptable or good circuit?

Monte Carlo runs

The number of runs depends on the accuracy you want to achieve on your yield estimates.
If you run 100 simulations and 2 of them are out of spec, you could - incorrectly - guess a yield of 98% and if you have 4 of the same designs in the same IC you could have a yield of 96%.
If you run 1000, you need 20 failures to get to the same result.
You need to choose the number of runs according to the confidence you want to achieve and - possibly - the time it takes to run the simulations.

Monte Carlo runs

In my opinion you should have zero failures.

Using 6 sigma design practice means it is highly unlikely to get a failure. 0.002ppm should fail (2 out of a billion). You are more likely to win the lottery than observe a failure.

Monte Carlo runs

I labored for years under a Monte-Carlo-centric regime
and I have a low opinion of the approach. For starters,
95% of the runs are worthless (in the sense that they are
not going to produce an outlier in any parameter of interest).
Yet your management will believe in this religion because
they want to believe, and you're stuck waiting out all those
worthless runs.

If you had the broader view and the time to spend and a
setup which is determinstic pseudorandom (as many are) you
can build a body of knowledge in short order, telling which
seeds are repeatably "bad actors". Then you could run them
first, a much shorter loop which is kind of like a "corners"
approach, but not restricted to the limited cases a digital
foundry might propose.

I have also found many times that MC statistics are sandbagged
to the point of being garbage, and you can get hit with sets
of params which would fail WAT. When I had the say-so,
I developed my own scripts which would skip iterations
that were shown to be WAT-rejectable.

Garbage in, waste 95% of your wall-time, garbage out. Yeah.

Of course somebody's CAD / modeling group -could- produce
a distribution- and detail-accurate statistics set, and your
design management -could- adopt more efficient approaches
to design criticism (such as I describe, or other). But MC
tools and approaches are sold specifically on the idea that
nobody has to think or be diligent - just push the button and
look at the histogram (much) later. It's thought, I guess,
to be a substitute for having designers who understand their
devices and their process.

I prefer not to work at places where this is thought to be a
good thing.

Monte Carlo runs

From my experience monte carlo simulation is very useful. We can consistently have 1st time success with no additional tweaks needed to reach production.

You have to have realistic models which should match silicon performance in the end. If not complain to your modeling group.

The idea is not to generate outliers, but to estimate production worthiness. Using the mean and standard deviation you should be able to determine that the design will yield well.

I mistake I have seen some colleagues make is running monte carlo with corner models. This should not be done as you will likely have some failures due to device parameters being way outside even worst case process windows.