Reiner Franke

Yes Giles, that are general consideration. I will guess that openbenchmarking.org is not a good source for machine selection for ngspice. Browsing little around and comparing the same CPU architecture Zen2, either on Notebook and Desktop there is a very huge difference which could be only because off unreliabilty boost clock operation in the tests. My guess is that this also happen to Zen3 for the 5800X versus 5700G

Nothing which make a difference. Openbenchmarking using a ngspice-34 release: https://openbenchmarking.org/innhold/8e889adb2b445383383fbfaedc9203060088aeed The only machine difference beside the clock speed in boost 4.4/4.6/5.3GHz is the L3 cache size with 16MB/32MB/32MB but the C2670 show a greater separation between 5600G and 7600 as the clock speed suggest.

The C7552 benchmark on openbenchmarking.org https://openbenchmarking.org/vs/Processor/AMD+Ryzen+5+5600G,AMD+Ryzen+5+5600X+6-Core,AMD+Ryzen+5+7600X+6-Core show the biggest discrepancy between Cezanne(APU) and Vermeer(Chiplet). The geometric mean difference over all test is -8.9% but in C7552 it is -50.2% Is there something wrong with the compile switches?

Is there some convergence experience which lead to: "KLU with --enable-klu, PIVTOL=1e-13" My guess is that with a higher PIVTOL setting a static elemination tree fails numerical. But anyway the result will be more numerical noise. As i pointed out above that static methods could fail if the conductance variation over time stick to the same pivots. This could change with models as less realistic ones could have conductance ratios which reduce finally numerical precision by this for static pivotin...

Hi Robert, i do not know the details about the integration of KLU into NGSPICE but i doubt that a non-modified usage of KLU works in all cases. I am out of simulator dev for 30y but could share what we know at this time from the research with Prof. Mathis. The modified nodal-approach generate time-varing conductances in transient evolution. This lead to pivot-changes over transient to allow numerical stability. If a static evaluation tree is used the pivot selection could numerical fail. So static...

Chandru, the error regarding the negative mobility i do not think relate to KLU because its just flaged before giving wrong result at all. My guess is that the KLU version also have different check levels basic physical model parameters or there are equations evalutions for the models which runs different on this KLU. I expect a modular option built for troubleshoot but i am not shure about your actual one. I am out of the simulator research community for now close to 30y and being surprised by the...

I want to contribute to a better understanding to separate ERC checks for manufacturing Solve checks to prevent wrong defined circuits with unclear outcome The problem i see is the floating nodes are connected by the simulator to some "useful" potential to enable solution. For the math standpoint voltage dependend capacitances obvious hide the dangerious "useful" potential which then lead to wrong AC results if the OP result is used for AC. In practice the next hiding level comes up if the circuit...

I got curious about the results of your findings. Reordering mean matrix, not netlist? Than it could be either that the Jacobi's leading to ill-condition, which could not be solved by full re-ordering, which i assume is done in the compile you use. The other failure is a possible incorrect temp equation which passed a check, or there is no check at all and the equations are not verified for all parameter space. What is the model which lead to the -40°C failure?