Geoffrey_Coram
|
jbdavid - Are you sure you've got it right? The reltol is applied to the signal itself, so the 1GV signal only has to converge to within 1MV (change in the value from iteration to iteration), but if you have normal 1.8V signals, they still have to converge to 1.8mV (unless you're using some funny tolerance option that uses the largest signal to set the tolerance for all nodes).
I think Ken's concern is with the other tolerance: vabstol. If you have something in the GV range, then "numerical noise" will appear around 1e-15 * 1e9 = 1e-6, which is the default for vabstol. It's easier to understand with currents: if you're adding 1GA and -1GA, you can't expect it to cancel better than 1e-6, so iabstol of 1pA is way too large.
|