< Tim Edwards> for some reason the `tt mm` corner does not s open-source-silicon.dev #sky130

<@U016EM8L91B> for some reason the `tt_mm` corner ...

Stefan Schippers

07/17/2022, 11:10 AM

@Tim Edwards for some reason the

tt_mm

corner does not show any mismatch in devices. Below

test_comparator.sch

schematic, part of the

xschem_sky130

test circuits does not show any variation. The waveform is actually an overlay of 40 runs, they are all identical. It used to work in the past, showing differences in the output waveform per each run. The netlist calls the

sky130.lib.spice lib

with

tt_mm

corner. Do I need to set some switches or parameters? Netlist attached as well.

test_comparator.spice

Krzysztof Herman

07/17/2022, 3:22 PM

@Stefan Schippers I am struggling with the same issue here, even in much simpler case of a voltage divider. Until now I've got any variations.

Krzysztof Herman

07/17/2022, 3:23 PM

VoltageDividerMM.spice,mm.txt

VoltageDividerMM.spice mm.txt

Tim Edwards

07/17/2022, 4:42 PM

@Stefan Schippers: Is the behavior different if you run ngspice multiple times instead of getting results from a

while

loop?

Tim Edwards

07/17/2022, 4:45 PM

I'm wondering if

reset

goes back to the same random seed due to the

option seed

command.

Tim Edwards

07/17/2022, 5:23 PM

I guess I should take at face value the comment that it used to work and assume that something broke.

Tim Edwards

07/17/2022, 5:29 PM

It is the parameter setting

MC_MM_SWITCH

that controls whether mismatch is turned on or not. It is always multiplied by the

agauss(.)

function, so that

MC_MM_SWITCH=1

randomizes the values and

MC_MM_SWITCH=0

does not. The

MC_MM_SWITCH

in device model parameter expressions is found throughout the files in

$PDK_ROOT/sky130A/libs.ref/sky130_fd_pr/spice/

. I don't see anything obviously wrong there, but I need to reinstall the PDK from scratch, since for open_pdks regressions can come from many different sources.

Krzysztof Herman

07/17/2022, 5:37 PM

Hi @Tim Edwards, I have took a look at the resistor model in the directory You've mentioned however haven't found the

MC_MM_SWITCH

. The only think which relates to the eventual mismatch modelling is this commented part. Is that right ?

Krzysztof Herman

07/17/2022, 5:37 PM

Copy code

* SKY130 Spice File.
* statistics {
*       mismatch {
*       }
*         process {
*               vary sky130_fd_pr__res_xhigh_po__var_mult dist=gauss std=0.025
*               vary sky130_fd_pr__res_high_po__var       dist=gauss std=0.025
*         }
* }

Tim Edwards

07/17/2022, 6:01 PM

@Krzysztof Herman: That's Spectre syntax. It's commented out because we didn't have time to figure out how to convert it to something ngspice-compatible at the time we generated the skywater-pdk repository. I commented out the lines but left them in so that the data would be available to make the conversion to ngspice syntax, which is what I'm doing in open_pdks. That said, it looks like the example you posted, for the xhigh_po and high_po resistors, does not get converted to an

agauss()

function anywhere. So that is probably what I need to track down.

Krzysztof Herman

07/17/2022, 6:03 PM

understood, thank You!

Krzysztof Herman

07/17/2022, 6:08 PM

I suppose the model definition should contain something similar to this line

Krzysztof Herman

07/17/2022, 6:08 PM

Copy code

toxe = {3.996598e-09+MC_MM_SWITCH*AGAUSS(0,1.0,1)*(4.148e-9*0.9635*(sky130_fd_pr__nfet_01v8__toxe_slope/sqrt(l*w*mult)))}

Christoph Weiser

07/17/2022, 7:34 PM

I don't know if this helps you, but bminch has made a video on this topic. I have modified my PDK succesfully with the script that he presents to run monte carlo simulations.

https://www.youtube.com/watch?v=fGxs2TnDgrU▾

Krzysztof Herman

07/17/2022, 7:35 PM

@Christoph Weiser, thanks I will give a try!

Tim Edwards

07/17/2022, 9:44 PM

Somebody please ping me on Tuesday. I want to take an in-depth look at the problem and get a fix out as soon as I can, but I'm traveling and won't be able to get around to it in the next two days.

Krzysztof Herman

07/17/2022, 9:45 PM

@Tim Edwards ok, I will remind You about the issue, have a nice trip!

Stefan Schippers

07/17/2022, 9:51 PM

Thanks @Tim Edwards for your investigations.

Stefan Schippers

07/17/2022, 9:55 PM

@Tim Edwards the while loop and the reset instruction does not reset the random number generator seed. In the test_comparator.sch i have (commented out) also an agauss() parameter for the VCC so each run would get a different VCC. I see VCC variations if i uncomment this agauss() parameter, so the MOS transistors not showing mismatches is probably due to some parameter switch, may be the MC_MM_SWITCH you mentioned.

Stefan Schippers

07/17/2022, 11:13 PM

@Tim Edwards looking into the spice model files it seems that

tt_mm

lib does set

MC_MM_SWITCH = 1

, so i don't know why we still see no mismatch variations in multiple runs. * Typical corner with mismatch (tt_mm) .lib

tt_mm

.param mc_mm_switch=1 .param mc_pr_switch=0 * MOSFET .include "corners/tt.spice" * Resistor/Capacitor .include "r+c/res_typical__cap_typical.spice" .include "r+c/res_typical__cap_typical__lin.spice" * Special cells .include "corners/tt/specialized_cells.spice" .endl

tt_mm

Setting a voltage source to value 'MC_MM_SWITCH' i see it has value '1' so the parameter is correctly set at simulation runtime.

Stefan Schippers

07/18/2022, 8:33 AM

@Tim Edwards I have a small script that starting from

sky130.lib.spice

displays content of all included files. Looking into this flattened big file there are

agauss()

functions only for resistors and for some sonos cells. so it is not surprising i see no variations. I see lot of

agauss()

parameters in file

.../libs.ref/sky130_fd_pr/spice/sky130_fd_pr.spice

which however does not seem to be included. May be we are in the middle of some big transition or change? Will update my open_pdks if you tell me to do so... Thanks

Tim Edwards

07/18/2022, 8:59 AM

@Stefan Schippers: There should not be such a file

sky130_fd_pr.spice

Tim Edwards

07/18/2022, 9:10 AM

@Stefan Schippers: What version of open_pdks do you have, and how was it built?

Luis Henrique Rodovalho

07/18/2022, 5:46 PM

There is a bug with monte carlo simulations. The simulator doesn't update the seeds between simulations that runs with less than 1 second. If you make another long simulation inside the loop, like a transient one, it should change the values.

Luis Henrique Rodovalho

07/18/2022, 5:48 PM

https://github.com/RTimothyEdwards/open_pdks/issues/201 See this thread. There are lots of repeated results, because there was not enough time between the simulations to change the seed.

Stefan Schippers

07/18/2022, 9:25 PM

@Tim Edwards I have built open_pdks from commit:

Copy code

commit 3a2a72e112d12771f528ed8b79b799222728e7eb (HEAD -> master, tag: 1.0.316, origin/open_pdks-1.0, origin/master, origin/HEAD)
Author: Tim Edwards <tim@opencircuitdesign.com>
Date:   Mon Jun 27 14:08:22 2022 -0400

doing the usual:

./configure --enable-sky130-pdk --prefix=/home/schippes

followed by

make && make install

. I have these files, both in the source tree and in the installed paths:

Copy code

schippes@mazinga:~/sda7/open_pdks$ locate sky130_fd_pr.spice
/mnt/sda7/home/schippes/open_pdks/sky130/sky130A/libs.ref/sky130_fd_pr/spice/sky130_fd_pr.spice
/mnt/sda7/home/schippes/open_pdks/sky130/sky130B/libs.ref/sky130_fd_pr/spice/sky130_fd_pr.spice
/mnt/sda7/home/schippes/share/pdk/sky130A/libs.ref/sky130_fd_pr/spice/sky130_fd_pr.spice
/mnt/sda7/home/schippes/share/pdk/sky130B/libs.ref/sky130_fd_pr/spice/sky130_fd_pr.spice

I think it's time to redo a new full install :-)

Tim Edwards

07/19/2022, 6:56 AM

@Stefan Schippers: I can confirm the presence of that file on doing a complete rebuild. My best guess is that some recent rework of the code in foundry_install.py has somehow caused it to believe that the "compile" option is always on. I'll debug it on my flight home today.

👍 1

Krzysztof Herman

07/19/2022, 9:13 PM

Hi, @Tim Edwards I hope You are back home, I send You reminder about the issue we have with mc_mm simulations, regards

29 Views

Open in Slack

Previous Next