Discrepency between Ngspice and Xyce simulation

[adrienluitot]

Environment

OS: Ubuntu 24.04
Tools: Qucs 24.4.1
Simulator: Ngspice-44 and Xyce 7.8

Expected Behavior

Having similar simulation results between Xyce and Ngspice.

Actual Behavior

We are designing a simple LNA at 2.45GHz. Here is its schematic and some ngspice simulation results:

On the right we see the Re(Z11) is 47.047Ω (with Ngspice).

When I run the same design, with the same parameters/variables, and the same simulation, but with Xyce, I get these results:

The discrepancy is not negligible here, mostly for Re(Z11), almost 10% difference.

For other results, there is not such a discrepancy, for instance, gain is pretty similar: 15.531dB for Ngspice and 15.561dB for Xyce.

I don't know if the error comes from:

• Me (Xyce installation maybe ?)
• Ideal components (maybe Xyce and Ngspice treat R,L,C slightly differently -> I will soon try with IHP's real components)
• Discrepancy between IHP's Ngspice and Xyce models
• Something else

Steps to Reproduce the Problem

1. Create a schematic in Qucs
2. Simulate with Ngspice
3. Copy whole schematic (design, parameters, simulations)
4. Adapt to Xyce (include & plugin)
5. Simulate with Xyce

[KrzysztofHerman]

@adrienluitot thank you for reporting this issue.
Since you use mosfets my first guess is that ngspice and Xyce use different Verilog-A compilers, openvaf and 'adms' respectively, in order to make PSP binary. Moreover you operate at relatively high frequencies, where the non quasi static PSP model part was disabled because of the openvaf compiler issues, what can bring us to this situation. If you wish you could investigate in this direction.

[adrienluitot]

@adrienluitot thank you for reporting this issue. Since you use mosfets my first guess is that ngspice and Xyce use different Verilog-A compilers, openvaf and 'adms' respectively, in order to make PSP binary. Moreover you operate at relatively high frequencies, where the non quasi static PSP model part was disabled because of the openvaf compiler issues, what can bring us to this situation. If you wish you could investigate in this direction.

Thanks for your reply! Indeed that may be the compiler, but I believe that the compilers should follow the models? So that it should rely more on the models rather than the compilers?

I don't know much about non quasi static, can you tell me more about it?
It would a pleasure to investigate, but I don't really know where to go. If you have any direction, that would be great!

[KrzysztofHerman]

@adrienluitot thank you for your interest. If you wish you could simplify the circuit and lower the frequency by a decade and then see the results. This way we could discard nqs. Our experience is that in general you will face differences between the simulation results using ngspice an Xyce.

[adrienluitot]

Sorry for the late answer. So I used the schematic given above and run the simulation with Ngspice and Xyce around 2.45GHz, 245MHz and 24.5MHz. And here is what I had:

2.45GHz:

245MHz:

24.5MHz:

Sure the discrepancy is more present (in percentage at least) at high frequency, but it seems also present at lower frequencies.