sdk@CS.UCLA.EDU (Scott D Kalter) (08/28/87)
We are having some difficulties with our simulation tools and would
appreciate any advice we can get.
Crystal:
It seems that the node resistance values that Magic extracts can be
grossly incorrect. The algorithm it uses (finding a rectangle from
the area and perimeter of the layer) is very bad with long thin
rectangles through which the signals travel across the thin side. For
example:
== <- Poly
==
-------X==X-------
| |X==X| |
|Pdiff|X==X|Pdiff|
| |X==X| |
| Con-|X==X| Con-|
| tact|X==X| tact|
| |X==X| |
-------X==X-------
^ ^
Pdiffusion
Magic will extract the resistance of the thin layer of Pdiffusion as
if the signal was flowing from top to bottom rather than from left to
right. So, if this diffusion was 1 lambda wide by 10 lambda tall then
the extracted resistance would be 10 times the sheet resistance of
pdiffusion rather than 0.1 times the sheet resistance. This is an
estimate that is 100 times larger than we want.
Crystal uses these resistance estimates in its algorithm for finding
the critical path of a circuit. Not only do the critical paths end up
having extremely pessimistic delays but it seems questionable whether
it has found the critical path at all.
We would like to know whether anyone has found a way to deal with this
problem. Hand-extraction seems like a particularly tedious and
error-prone solution, especially on very large circuits.
Spice:
We are interested in what spice level 2 parameters people are using
for the fabrication through MOSIS service. Why are you using those
particular parameters and where did you get them?
Thanks for your help,
Scott Kalter <sdk@cs.ucla.edu>
Jae Cho <jcho@cs.ucla.edu>gwu@vlsi.cs.cmu.edu (George Wu) (09/04/87)
There are several other parasitic circuit extractors out there which try
to handle resistance better than Magic's extractor. At least a few which I
know of are:
1) RC from Brish, Pardo, and Ben-zvi at Motorola Israel. This looks
interesting since it has some interesting methods which don't involve
finite elements. Primarily looks at edges to determine the direction
of current flow.
2) EXCL by McCormick at MIT. Here accuracy seems to have received
the highest priority, since finite elements is used. However, there
are some optimizations used to help cut down the extraction time.
3) FRED by Mike Trick here at C-MU. Again, another fast(er) extractor
which tries to find a balance of both speed and accuracy. Edges
are used to determine current flow direction. Unfortunately, FRED is
for NMOS only. Mike and I are working on Paratice, which will be a
technology independent extractor based on FRED. It should be ready by
the end of the month.
Also, I believe there is work being done at Berkeley on a new extractor which
is much more intelligent about parasitics. I think they published a paper on
it, though I'm not familiar with their work and can't really say much about
it.
As for the Spice level two model parameters, below are the parameters I
use. they were given to me by a professor when I took is course last year. I'm
not sure where he got them from.
.MODEL CMOSN NMOS LEVEL=2.00000 LD=0.280000U TOX=530.000E-10
+NSUB=1.000000E+16 VTO=0.864362 KP=3.310097E-05 GAMMA=1.53162
+PHI=0.600000 UO=230.648 UEXP=1.001000E-03 UCRIT=999000.
+DELTA=1.23086 VMAX=65794.7 XJ=0.400000U LAMBDA=1.554857E-02
+NFS=9.977800E+11 NEFF=1.001000E-02 NSS=0.000000E+00 TPG=1.00000
+RSH=20 CG.MODEL CMOSP PMOS LEVEL=2.00000 LD=0.280000U TOX=530.000E-10
+NSUB=1.365450E+15 VTO=-0.725007 KP=1.246790E-05 GAMMA=0.628686
+PHI=0.600000 UO=100.000 UEXP=0.138792 UCRIT=13550.7
+DELTA=2.11805 VMAX=100000 XJ=0.400000U LAMBDA=4.345936E-02
+NFS=1.038954E+12 NEFF=1.001000E-02 NSS=0.000000E+00 TPG=-1.00000
+RSH=55 CGSO=4E-10 CGDO=4E-10 CJ=3.6E-4 MJ=0.5 CJSW=6.0E-10 MJSW=0.33
I hope all this helps.
George