[SOLVED] Monte Carlo Simulation Without Statistical Data from foundry

[rmanalo]

Hello everyone,

For various reasons, our laboratory only has the library file for the 0.18um TSMC technology node. Some sort of "hand-me-down" from an instructor. I'd like to perform a monte carlo simulation using the HSPICE ver: A-2008.03. Please give feedback if what I'm doing is close to what is expected.

From the library file,

.LIB TT
.param
+toxn   = 4.08E-09        toxp   = 4.08E-09       
+dvthn  = 0               dvthp  = 0              
+dxl    = 0               dxw    = 0              
+cjn    = 0.001000266     cjp    = 0.001121       
+cjswn  = 2.040547E-10    cjswp  = 2.481E-10      
+cjswgn = 3.340547E-10    cjswgp = 4.221E-10      
+cgon   = 3.665E-10       cgop   = 3.28E-10      
+hdifn  = 2E-07           hdifp  = 2E-07          
.lib 'mm018.l' MOS
.ENDL TT
***************** CORNER_LIB OF SNSP MODEL ***************
.LIB SS
.param
+toxn   = 4.213E-09       toxp   = 4.213E-09      
+dvthn  = 0.1             dvthp  = -0.067         
+dxl    = 1.33E-08        dxw    = -2.2E-08      
+cjn    = 0.001050279     cjp    = 0.00117705     
+cjswn  = 2.142574E-10    cjswp  = 2.60505E-10    
+cjswgn = 3.507574E-10    cjswgp = 4.43205E-10    
+cgon   = 3.48175E-10     cgop   = 3.116E-10    
+hdifn  = 2E-07           hdifp  = 2E-07          
.lib 'mm018.l' MOS
.ENDL SS
***************** CORNER_LIB OF FNFP MODEL ***************
.LIB FF
.param
+toxn   = 3.947E-09       toxp   = 3.947E-09      
+dvthn  = -0.1            dvthp  = 0.067          
+dxl    = -1.33E-08       dxw    = 2.2E-08       
+cjn    = 0.0009502527    cjp    = 0.00106495     
+cjswn  = 1.93852E-10     cjswp  = 2.35695E-10    
+cjswgn = 3.17352E-10     cjswgp = 4.00995E-10    
+cgon   = 3.84825E-10     cgop   = 3.444E-10    
+hdifn  = 2E-07           hdifp  = 2E-07          
.lib 'mm018.l' MOS
.ENDL FF
***************** CORNER_LIB OF SNFP MODEL ***************
.LIB SF
.param
+toxn   = 4.08E-09        toxp   = 4.08E-09       
+dvthn  = 0.1             dvthp  = 0.067          
+dxl    = 0               dxw    = 0              
+cjn    = 0.001050279     cjp    = 0.00106495     
+cjswn  = 2.142574E-10    cjswp  = 2.35695E-10    
+cjswgn = 3.507574E-10    cjswgp = 4.00995E-10    
+cgon   = 3.665E-10       cgop   = 3.28E-10      
+hdifn  = 2E-07           hdifp  = 2E-07          
.lib 'mm018.l' MOS
.ENDL SF
***************** CORNER_LIB OF FNSP MODEL ***************
.LIB FS
.param
+toxn   = 4.08E-09        toxp   = 4.08E-09       
+dvthn  = -0.1            dvthp  = -0.067         
+dxl    = 0               dxw    = 0              
+cjn    = 0.0009502527    cjp    = 0.00117705     
+cjswn  = 1.93852E-10     cjswp  = 2.60505E-10    
+cjswgn = 3.17352E-10     cjswgp = 4.43205E-10    
+cgon   = 3.665E-10       cgop   = 3.28E-10      
+hdifn  = 2E-07           hdifp  = 2E-07          
.lib 'mm018.l' MOS
.ENDL FS

What I intend to do is use the HSPICE variation block to simulate and vary each parameter by applying gaussian distribution to each where the 3-sigma deviation is the SS or FF value for that parameter. For example the vth should look like.


Other parameters are also done in this manner. My goal is to produce a statistical histogram for the D2D or interdie variation.

Questions:
[1] Is this method a good alternative for simulation without the statistical data from the PDK?
[2] does mismatch have a significant effect in D2D variation? (since I did not include it)

 

[pancho_hideboo]

My goal is to produce a statistical histogram for the D2D or interdie variation.

What do you want to mean by "D2D" ?

[1] Is this method a good alternative for simulation without the statistical data from the PDK?

Basically, it is OK.
However abs(FF-TT) is not always equal to abs(SS-TT) generally.

I set 6*Sigama=abs(FF-SS). Mean=TT.
This is a little lllegal.

[2] does mismatch have a significant effect in D2D variation? (since I did not include it)

Do you mean "Mismatch Variation" and "Process Variation" ?

FF, TT, SS are "Process Variation".

"Mismatch Variation" is called as "Local Variation" in HSPICE.
"Process Variation" is called as "Global Variation" in HSPICE.

"Mismatch Variation" is dependent on device size.

 

[ThisIsNotSam]

generally speaking, you are on the right track. the problem is that foundry models are smart, and they have interdependence built in. Say, if parameter A goes up, B goes down. In your approach, you would be assuming A and B can both go up. The easiest example of this is related to doping parameters, that favor either N or P sides, but never both at the same time.

 

[dick_freebird]

You can do poor-boy process stats by looking at the foundry
wafer acceptance test (WAT) limits, call center-min and
center-max "three sigma" (although you do not know a priori
whether the distribution is Gaussian or truncated mystery
shape). This might get you at least u0, tox, VT but you may
or may not be getting Leff / deltaL (critical at min L), deltaW,
variation in overlap capacitances (spacer and implant variation)
and so on. And only a more comprehensive production yield
/ yield enhancement report would likely give you any insights
to cross-correlation (e.g. VTN/VTP have common and individual
influences; tox ought to be common; deltaL has common
spacer but separate implant influences, etc.).

Obtaining the deeper data would probably require a good working
relationship and legalities. Or, access to the already-done PDK
to translate to your preferred tools set, whatever that is. The
WAT limits, you ought to be provided before you sign up for
a wafer lot (but how far from that, if ever, are you?).

Simple is how we used to roll, before we had whole departments
for technology development and PDK / CAD fit-up. You can get
it done. But you can't convince any minor middle manager that
this is a can't-fail proposition, and they'll take a knee and go
with the more expensive effort as long as it's anybody else's
dime.

 

[BigBoss]

Those are not Process Parameters, they are Device Parameters of the transistor model that are relied in anyhow to Process Parameters.
Since you don't know the Process Parameters ( and their statistical variations ) there is no chance to do MC simulation.Process Parameters are measured by Foundry guys on many test wafers and then added to PDK.You cannot do that..

 

[dick_freebird]

Sure, you can. Because I have multiple times. As good as a
10-man PDK development group and another handful of modeling
dudes? No. Good enough for the first few product designs on
the processes I was first-in on? Yeah, boy.

There's not a bean's worth of difference between the ratiometric
scatter on WAT VTlin, and PDK VT0. WAT ETOX vs PDK TOX, ditto.
You don't need data points, you just need mean and sigma so you
can figure =sigma/mean and apply to the model stats.


From historical WAT data you can get scatter plots which will tell
what params have a correlation (if you care to grind it out and
inspect it all). You may not get that data without some heavy
petting, but you might ask.

There's foundries and there's foundries. I've worked at places
where you could not design on anything that didn't have the
full up fancy multiple-lots-worth-of-statistics PDK fully blessed,
and I've also designed products before the process was even
frozen, because nobody wanted to wait, at the same company.
Rules are rules until somebody smells money.

But you could also save yourself a lot of effort and just do a
corners model set based on WAT limits, if you don't care a lot
about mismatch stats (and critically, their relation to device
geometry).