It is probably a feature of your circuit rather than a problem with the simulator. If you sweep temperature in the opposite direction, you will probably see classic hysteresis behavior.

-Ken

When a circuit exhibits a jump like you are describing it is because it has solution curve that has folds in it, as shown with the solid line below.



Here the horizontal axis represents temperature and the vertical axis represents a signal value, say the output voltage.

Starting from the left, the temperature increases until it reaches the lower fold, at which point the output voltage jumps discontinuously. In this case it jumps up to the upper part of the curve (dashed upward facing arrow). If the temperature then began to decrease, the output voltage would stay on the upper part of the curve until it reached the upper fold, at which point it would jump down to the lower part of the curve (dashed downward facing arrow). In this way, the voltage/temperature curve would exhibit classic hysteresis.

If you started at a point between the two folds, then going up you would see a jump, but going down you would not because the upper fold occurs at a temperature less than -40C.

You are almost correct in your belief that DC sweeps are memoryless. To speed up DC sweeps, Spectre extrapolates from the previous solution point to the next. It then starts the DC iteration from the extrapolated point. This tends to cause Spectre to faithfully follow the curve as long as it can. It is this extrapolation that represents the memory in the DC sweep.

-Ken

Yes, from what you have told me, I believe your circuit has two stable and one unstable operating points at temperatures below where the jump occurs.

-Ken

Hi, Ken,

I did not find the jump when I sweep temperature from 125 to -100.

But if I change gmin from 1e-12 to 1e-15, there is no jump from either direction. Does this mean the circuit has no mutiple solution problem?

-driveforce

driveforce,

  Are all of the devices in circuit properly biased? In the past, we
have seen issues when the gmin loaded nodes because the transistors
were all biased off. In our case, by increasing the tolerance of gmin,
the effect of device leakage currents became larger then the effect of
gmin and the circuit setup properly. After we corrected the underlying
circuit issue, the default gmin was sufficient.

                                                                        Best Regards,

                                                                              Sheldon

Hi, driveforce,

A bandgap reference has more than one steady state. So a start-up circuitry is required to make the bandgap reference work in the desired point. Many textbooks describe this feature of a bandgap circuit.

I think your start-up circuitry doesn't work at very low temperature. It is easy to correct it.


Best regards,

Yawei