Very slow Cace simulation on ulpcomp2 input offset example
azwefabless opened this issue · 5 comments
cell is ~/cheetah_v3_analog/dependencies/sky130_icrg_ip__ulpcomp/
Run time is exceptionally long for me but about 100x faster for Tim. Suspect an IO setting difference that Tim has set but I have not.
The simulation with the speed problem is input offset voltage
Hi, thanks for the bug report!
The repository for sky130_icrg_ip__ulpcomp is here: https://github.com/efabless/sky130_icrg_ip__ulpcomp/
And the problematic parameter is that one here, right? https://github.com/efabless/sky130_icrg_ip__ulpcomp/blob/e0244073f7b21e02c1f6d787516cf0ae9cfe5436/cace/sky130_icrg_ip__ulpcomp2.txt#L425
I've seen something similar with a transient simulation as well. Are you getting a lot of "Reference value" output?
Are you running CACE via GUI or CLI? If you are working via cli, could you try passing --sequential? Please make sure cace is up to date as --sequential was broken previously.
If you are working via gui, you can enable sequential mode in the settings.
Yes, Correct repo and correct file. There are 2 sims that are dying in what appears to be similar manner. Both output enormous amounts of Reference value data to the GUI. Tim runs the same config file and it finishes in minutes for him. Mine is still going after 20 minutes with 8 railed CPU cores. Sequential might be a good workaround but is not a fix. I will try it today.
Thanks! I just ran it on my side (fresh pull) and here it seems to work well:
cace --source schematic cace/sky130_icrg_ip__ulpcomp2.txt -p offset_error
I believe it could be a tooling issue. I am using Nix to set up my tools, it uses the stable version of ngspice.
There's no installation documentation on that yet, but it's basically the same as in OL2: https://openlane2.readthedocs.io/en/latest/getting_started/common/nix_installation/index.html
My PDK is bdc9412b3e468c102d01b7cf6337be06ec6e9c9a (2024.01.10) (enabled).
Which tool and PDK versions do you use?
Taking a second look at it, it doesn't seem to work well ^^
The results are nonsense, also I get measurement errors from ngspice:
Error: measure vhigh find(AT) : out of interval
meas tran vhigh find v(inp) when v(vout)=1.65 cross=1 failed!
Is the repository in a working state? I am asking because I had to replace some /home/ttuser/... paths with {PDK_ROOT}.
I'm having the issue after swithing to a prereease version of the PDK that Tim requested. Also note that there are a lot of convergence related comments in the terminal. I suspect either a models issue or an environment setting variable around convergence or ngspice setup that is missing from the default config information.