Noise goes wrong with psp103_nqs

[dwarning]

This setup was simulated correct with model compiled by the old openvaf Version (noise activated) from Semimod download page

Using openvaf from origin/branches/osdi_0.3_noise_mfactor under linux (BTW - not easy to pick-up the right branch, worry!) I get:

Same behaviour under windows with your provided binary.

Used psp103_nqs.va model Version 103.8.2

I have not the problem with older psp103 version 103.6.0 from same repository. Model w/o noise problem

[dwarning]

I isolated the problem: Commit 1529801 - the replacement of Geoffrey's 477e71c seems the culprit. Interesting because with 477e71c we can not compile psphv.va - see our discussion in 137.
The shown testcase has no multiplier.

[arpadbuermen]

I observed the same behavior with
Version: 103.6.0 (PSP), 200.5.0 (JUNCAP), December 2017
Can you recheck.

The problem originates from noise source
I(NOII) <+ white_noise((nt / mig), "igig");
When this source is commented out, the problem goes away.

If I leave the line below unchanged
I(NOII) <+ white_noise((nt / mig), "igig");
and comment out
I(DI,SI) <+ sigVds * sqrt(MULT_i) * migid * I(NOII);
The problem also disappears.
It seems the source of the problem is referring to a branch current of a branch with only a noise contribution.

A simulation in some reference simulator that supports Verilog-A would be welcome. Just to find out, what the correct behavior should be.

[arpadbuermen]

It seems that the code for computing derivatives is incorrect. In the special case when you contribute a current to a branch that has only a noise component

electrical A, B;
branch(A) br1;
I(br1) <+ white_noise(...);

and then use that same current to contribute to the current of another branch

I(B) <+ I(br1);

a new unknown for I(br1) is created. Now who does that kind of exotic stuff in Verilog-A? Turns out the PSP people do :) All hell breaks loose because the derivative dI(B)/dI(br1) is incorrectly computed.

The derivative is computed correctly if I(br1) has a small resistive contribution, i.e.

I(br1) <+ 1e-100+V(br1) + white_noise(...);

The workaround until the bug is fixed is to change the file PSP103_module.include. Change

I(NOII)      <+  white_noise((nt / mig), "igig");

into

I(NOII)      <+  1e-100*V(NOII) + white_noise((nt / mig), "igig");

The bug seems to have floated up to the surface when correct $mfactor handling was added.

Please confirm if it works for you.

[dwarning]

I can follow you analysis and confirm the workaround by adding a tiny resistive component.

If I copy over the file openvaf/sim_back/src/dae/builder.rs from @gjcoram the problem disappears (but crash in psphv.va remains). I thought I can locate the root cause by inspection the two files but your modifications in this file are to extensive an my rs skills are very limited.

[arpadbuermen]

Seems to be closed. Build OpenVAF and try it out before I close the issue.

[dwarning]

Can confirm, the psp noise bug is fixed. Thank you.
BTW: psphv can still be compiled as before.