Monte Carlo simulation with verilog-a model using Spectre

There's an example in this solution of how to do this.

Andrew.

Hi Andrew,

Thank you for the link to the solution. I've tried it myself and it works. However, I wonder if it's possible to define the statistical parameter with respect to another parameter of the model.

Say I have modeled a current source (see below) and I want the standard deviation to be a certain percentage of the value that the user defines in the schematic for the specific model. 

--

// VerilogA for user, vccs_mc, veriloga

`include "constants.vams"
`include "disciplines.vams"

module vccs_mc(vpos,vneg,ipos,ineg);

inout vpos,vneg,ipos,ineg;
electrical vpos,vneg,ipos,ineg;

(*cds_inherited_parameter*) parameter real gm_mc = 0;

parameter real gm = 1u;
localparam real gm_effective = (1+gm_mc)*gm; // nominal transconductance plus monte-carlo mismatch effect

analog begin
I(ipos,ineg) <+ gm_effective*V(vpos,vneg);
end

endmodule

//

How shall I define the variation of the parameter gm_mc in the .scs file such that is taken into account properly ? Currently I have the following:

parameters gm_mc = 0

statistics{

mismatch {

vary gm_mc dist=gauss std=1 percent=yes

}

}

But when I run the Monte Carlo the simulator complains there's no statistical variation.

Thanks in advance.

Lucho

If I run with your statistics block, I get:

Warning from spectre during Monte Carlo analysis `mc'.
WARNING (SFE-3224): Variance of the mismatch parameter `gm_mc' is evaluted to be zero.
Error found by spectre during Monte Carlo analysis `mc'.
ERROR (SFE-3225): Variance (or standard deviation) of statistical
parameters cannot be zero! Set `ignorezerovar=yes' in global options to
bypass this check.

This is because you have percent=yes and so the standard deviation is set to be 1% of the mean value (which is 0). That's not going to work.

There's no need to make the standard deviation a function of an instance parameter. What you would do is have a normalized statistical parameter - e.g. a mean of 0, and standard deviation of 1, and then scale it in the model. In this case your model has a standard deviation of gm (if I remove the percent=yes) because you're multiplying gm_mc by gm. If you wanted it to be 1% of gm, you could write:

localparam real gm_effective = (1+gm_mc*0.01)*gm; // nominal transconductance plus monte-carlo mismatch effect

Regards,

Andrew.

Thanks for the explanation and the workaround Andrew. Your approach indeed does the trick.

Regards,

Lucho.