Problem about Monte Carlo analysis

Hi,
I have some problem with Monte Carlo analysis.
Each time I ran the Monte Carlo analysis, I got the same
results. It seems that the seed of the random function is always the same.
Is there some way to deal with it? Or I miss some options?

Thanks.

 

I don't see a field to specify a seed in the MC GUI.

The documentation says:
If you do not specify a seed, then each time that you run the
analysis, you get different results; that is, a different stream of
pseudorandom numbers is generated. If you do not specify a seed, the
Spectre simulator uses the Spectre process id (PID) as a seed.

I don't see this to be the case however. If I do a Simulation->Create
Input Files I always get seed=1 in the MC section of the netlist.

 

Yes, the seed is currently fixed. There's a PCR to allow the seed to be
specified in the UI (I don't remember the PCR number) - I think it has
come up in the past on this group (check google).

The idea behind having a fixed seed was to allow you to have reproducable
results - and you can (sort of) workaround the problem by changing the starting
run number on the form.

I'm assuming Yan that you're not saying that the results for each point in
the monte-carlo are the same as all the other points, but rather that the
results for the whole set of runs stays the same?
If the results are the same point to point, then that would suggest that
you either have models which don't vary, or more likely that you have
expressions for your outputs which reference the wrong set of results.

Andrew.

 

FYI, the PCR number is 432223.
---
Erik

 

Indeed,

it is very useful to be able to go from a given MC run in
results/waveform back to the state of the process and models. Around
here, the typical designer has already pushed the envelope on disk usage
and CPU time before he starts to do MC runs, so he will not be able to
save all results, and he will not have the time to simply rerun the
first 73 runs when he notices something "interesting" in the 74th MC run
and want to know more about it.

Even with the possibility to set the seed to the value of a given MC
run, there is still quite some work to relate the seed value to the
process parameter and model parameter values. And the process params are
what you want to know so that you can understand how the process state
makes the circuit fail, and be able to correct.
Since MC, corner runs and such are late in the electrical design
phase, at a time when people a also busy documenting, preparing
presentations, meetings, and close to a deadline, they will almost never
go through much trouble to trace back what this particular MC run was
doing to their circuit. They will rather make assumptions (that they don
t check by simulation), fix the circuit, start a new set of simulations
and check only that the worst case situation doesn happen anymore.