The workarounds I found to this limitation are
1) using different model files for the different device sizes, or
2) using subcircuits instead of primitives for the devices. In the subcircuit you can compute the mismatch parameters according to the dimensions and then call the model with the right parameter values. This requires creating subcircuit definitions for your devices, so I hope you have some CAD people who can handle this.

Paul

Paul,

Thanks for your advice. I tried the subcircuit method, it works fine with one problem:

The multiplicity factor m. m is not accessable in subcircuits. The result is that standard variation of m devices in parallel generates m times of variation rather than sqrt(m).

Any comments?

Paul,

Thank you very much for your help. I am using Spectre and modifying a TSMC Sige BiCMOS design kit.

Now I can do Monte Carlo simulation on most devices. However, It seems that there is an ultimate barrier:

The Monte Carlo simulation will fail if model binning is used. In my case, MOS devices are modeled in many bins.

I have no clue on how to work around this problem.

The problem was:

When I used monte carlo simulation to simulate CMOS devices in sub-circuits which instanciate a MOS device with model binning, the first run is fine. But the second run fails with a message citing "segmentation fault".

It is fine if I do not use model binning.

> Spectre also provides a non-Monte Carlo based mismatch analysis. Have you tried it?

Mighty Mouse,

I am not aware of this analysis. I am mostly using ELDO, but I would be interested in this Spectre feature. What Analog Environment menu can I find it in?

For the binning problem, I don't have any solution, I used this with non-binned models (older technology). I agree that you should try to get support from Cadence for this bug. BTW, there is a third possibility I didn't mention in my first answers. If you have an EKV model (instead of BSIM), this model contains the mismatch parameters in the model, so there's no need to go the subcircuit way... Unfortunately, EKV doesn't seem to become an foundry standard.

Paul

You have to place the expressions into every model section of your binned models. Here is an example of how to do this with binned modls:

parameters vth0mis_n=0.0
parameters vth0mis_p=0.0
parameters u0mis_n=0.0
parameters u0mis_p=0.0

statistics{
 process{
   vary vth0mis_n dist=gauss std=0
   vary vth0mis_p dist=gauss std=0
   vary u0mis_n dist=gauss std=0
   vary u0mis_p dist=gauss std=0
 }
 mismatch{
   vary vth0mis_n dist=gauss std=4e-3
   vary vth0mis_p dist=gauss std=5e-3
   vary u0mis_n dist=gauss std=0.4e-2
   vary u0mis_p dist=gauss std=0.6e-2
 }
}

inline subckt nmos (d g s b)
        parameters w=1u l=1u wd=0.6e-6 ws=0.6e-6 ad=wd*w as=ws*w pd=(2*wd) + (2*w) ps=(2*ws) + (2*w)
        nmos (d g s b) nmos w=w l=l ad=ad as=as pd=pd ps=ps


model nmos bsim3v3  {
    1: type=n minr=.... lmin=.... lmax=.... wmin=..... wmax=.....  
+ ............    vth0=0.5 + dvthn + vth0mis_n/sqrt(w*l*1e12)    ..............    
+ ......         u0=0.04*(1+u0mis_n/sqrt(w*l*1e12))   ......................

   2:    // same way

}  

Eugene

[quote author=EVI  link=1066706959/0#11 date=1070292390]

inline subckt nmos (d g s b)
        parameters w=1u l=1u wd=0.6e-6 ws=0.6e-6 ad=wd*w as=ws*w pd=(2*wd) + (2*w) ps=(2*ws) + (2*w)
        nmos (d g s b) nmos w=w l=l ad=ad as=as pd=pd ps=ps

[/quote]

Can you explain the above line is doing? Do you include only once or you need for every MOS