As in, new version? Or just in general?
Computational solutions are amazingly difficult to realize. As powerful as SPICE is, there are still a multitude of "correct", reasonably sized circuits, which will not simulate in a reasonable time frame, for any environment parameters.
The best approach is to minimize unstable, poorly defined nodes, and get as many continuous derivatives as possible. Examples of bad circuits: capacitors in series (degenerate voltages -- if Rshunt = inf, gives "singular matrix" error), inductors in parallel (same thing, but for Rseries = 0, and currents being undefined), behavioral blocks using discontinuous (e.g., IF, SGN) or piecewise-continuous (TABLE, LIMIT, etc.) functions, circuits that are highly nonlinear (e.g., oscillators, switchers, discrete (usually non-behavioral*) logic) that don't also have realistic parasitics included in the model (ESR, ESL, strays, etc.).
*Behavioral logic runs in a different subsystem, and only interfaces with analog simulation at port pins. Which have their own range of concerns, so I'll just leave that out for now..
Of course, rich, relatively accurate models (those including most nonlinearities, parasitics, time and state dependencies, etc.) tend to run very slowly, for the fact that they have a lot to do... but they also tend to go more smoothly. At least... in as much as, when they don't get randomly hung up on a small timestep, and judging performance relative to the number of nodes in the model.
Tim