Using Spectre MDL to optimize transistor multiplicity (integer parameters)

Dear ConradJ,

ConradJ said:

Is it possible to specify an optimization parameter as an integer or achieve a similar effect?

I am neither an SDL expert nor a regular MDL user, but the thought that came to mind as I don't know of a way to force the optimizer to perform an integer search, is to utilize a couple of foreach loops for sets of integer values that are realistic.

For example, suppose  the variable M_nmos is the nmos multiplier and M_pmos is the pmos multiplier. Suppose you expect the multiplier values to be between 5 and 10 for M_nmos and 10 and 20 for M_pmos. You may have other parameters that are varying (such as channel length for example) in the optimization. You might consider the following basic syntax:

Shawn

Not sure how Shawn's suggestion is going to work since that just adds two loops around the search - and so will just do 66 optimisations, rather than one optimisation to find the two M factors.

I think it should work (based on a previous change request) by simply declaring the parameter in the MDL file (above the mvarsearch) as an integer:

int M_nmos=5
int M_pmos=10

(you have to declare an initial value when declaring it).

I've not got a decent test case to try this out on to hand. If I have a moment, I'll do so later today.

Andrew

ConradJI spent a bit of time trying to get this working. First of all I found that making a spectre parameter an integer seemed to prevent it getting passed to the netlist, so instead I tried this:

//  mvar.mdl
//  (C) 2008 Cadence Design Systems, Inc.
//
// This example is to find optimal values of design parameters 
// by running a multi-parameter, multi-goal search function
//

// alias measurement definition
alias measurement trans {

  // definition of transient analysis with autostop turned on
  run tran( stop=1u, autostop='yes )

     // computation of risetime for signal V(out) from 10% to 90%
     export real rise=risetime(sig=V(out), initval=0.5, \
                      finalval=3.0)

     //computation of falltime for signal V(out) from 90% to 10%
     export real fall=falltime(sig=V(out), initval=3.0, \
                      finalval=0.5)

}

// Print results to a specified file beginning with "to=" keyword
print fmt ("\n****Intermediate Results of Optimization Searching****\n\n") to="mvarint.print"

// continue to print in user-format with "addto=" keyword
print fmt ("%-15s%-15s%-15s%-15s\n", "mp","mn","rise","fall") addto="mvarint.print"

int mpint=1
int mnint=1

// definition of a mvarsearch function
mvarsearch {

         // definition of multiple optimization goals
   	 option {
   	    accuracy = 1e-3     // convergence tolerance of trans
   	    deltax = 1e-3       // step length
   	    maxiter = 100       // limit to 100 iterations
   	 }

         //  definition of multiple parameters to be optimized
   	 parameter {
   	    {mpint, 2, 1, 10} //init_val, lower_val, upper_value
   	    {mnint, 2, 1, 10} //init_val, lower_val, upper_value 
   	 }

         // execution of the measurement alias
   	 exec {
            mp=mpint
            mn=mnint
            print fmt("%-15e %-15e", mp, mn) addto="mvarint.print"
   	    run trans
            print fmt("%-15e%-15e\n", trans->rise,trans->fall) \
                      addto="mvarint.print"
            run instParams
   	 }

         // definition of goal values to be minimized
   	 zero {
   	    tmp1 = trans->rise-25p
   	    tmp2 = trans->fall-25p 
   	 }
}

I made the parameters in the netlist mp and mn, and then optimised mpint and mnint. This seems to transfer to the netlist OK, but something really isn't working when I try this (I'm not sure the optimisation loop is correctly handling discontinuous variables).

I don't have the bandwidth to test this more (I have vacation coming up) so I suggest you contact customer support.

Andrew

Thank you for your response! I agree the optimizer is having trouble with the discontinuities. I will contact support and let you know what I find. Enjoy your vacation!

Dear Andrew,

Andrew Beckett said:

Not sure how Shawn's suggestion is going to work since that just adds two loops around the search - and so will just do 66 optimisations, rather than one optimisation to find the two M factors.

Indeed, this will perform multiple optimizations. However, the concept is that it will produce a table with the results of each optimization. From inspection of the table, one can select the specific case that provide a global optimum.

Although this methodology does not produce a single optimum result, it was my attempt to provide an optimal solution where the optimization algorithm does not support an integer-only constraint for the parameter to be optimized.

It might be an alternative that may be considered pending any customer support delays.

Shawn

Dear Andrew,

Andrew Beckett said:

Not sure how Shawn's suggestion is going to work since that just adds two loops around the search - and so will just do 66 optimisations, rather than one optimisation to find the two M factors.

Indeed, this will perform multiple optimizations. However, the concept is that it will produce a table with the results of each optimization. From inspection of the table, one can select the specific case that provide a global optimum.

Although this methodology does not produce a single optimum result, it was my attempt to provide an optimal solution where the optimization algorithm does not support an integer-only constraint for the parameter to be optimized.

It might be an alternative that may be considered pending any customer support delays.

Shawn