Hello,

I have come across an issue that I have not seen addressed.

I am simulating an oscillator in spectre. I would like to simulate it for 100 us. For the first 50 us I would like moderate accuracy. For the remaining 50 us I would like conservative accuracy.

How do I do this?

This really comes down to an issue of speeding up the simulation time. I see changing the accuracy as the main means, but there are some choices:

1. Find out how to run portions of the transient in moderate and conservative. I have used writefinal and readic, but that does not work as I desire -- if at all. There are many errors and jacobian matrix issues. How about restart or checkpoints?

2. Disable writing of data for a portion of the simulation (through the outputstart hold-off time). This knocks time off the simulation, but I feel there must be a better way -- especially if the change of accuracy solution saves even more time.

3.  ??

I have contacted Cadence tech support and they are currently pondering this issue.

Any ideas or suggestions? If anyone else has any related tips on how to run oscillators quickly and accurately, please speak up. I have thought about periodic steady-state (pss), but that still requires the transient to run.

-CC

P.S. If I hear a solution from Cadence, I will post it here.

The writefinal/readic approach will work, you just have to be careful with your sources. Consider the case where there are two sources driving your circuit, one that produces a step that stabilizes after 100 ns and the other that produces a sinewave with a 300ns period. Now assume that you run to 1us and write the state of the circuit to a file. If you now rerun the simulation from t=0 and use readic to start the circuit in the previously saved state, you will have problems because you are trying to force the circuit into an unnatural state because the sources at t=0 produce different values than they did at t=1us.

There are two ways to address this. One, you can simply set the start time to t=1us, as you found out. Alternately, you can adjust the sources on subsequent runs so there initial value corresponds the their value when you recorded the state.

Another approach to increasing accuracy at a particular point in time is to include a Verilog-A module whose only purpose is to shrink the time step. This is not the same as tightening the tolerances, but can be quite effective as long as your timesteps do not vary widely over the course of a single cycle (if they do, then forcing small time steps will still work but can be very expensive). To do so, use something like this ...

module limit_timestep();
parameter real tstart;
parameter real maxstep;
analog begin
 if ($abstime > tstart) begin
   $bound_step(maxstep);
 end
end
endmodule
In Spectre you can get the same effect by adding a disconnected sinusoidal source into your circuit that starts producing a high frequency sine wave at the point where you want more accuracy. Don't use a pulse, as it causes the simulator to use backward euler integration at the sharp corners, which tends to degrade accuracy. Given a choise, the Verilog-A approach is the best as it is the most direct and gives the finest degree of control.

Be aware that this idea of running for a while then tightening tolerances may not give you the accuracy you seek if your circuit exhibits long time constants (which is often the reason for the long initial simulation). The reason is that the error in the results that exists at the end of the long initial simulation interval will take a long time to decay because of the long time constants.

-August West