It works provided you are using bsim3v3 or bsim4 models, and you have at least 512 instances (I think that's the number) of these devices.

Some time ago I did some benchmarks and got some noticeable improvement between 2 and 4 CPUs, but not anything significant above 4 CPUs. Even with 4 CPUs the processors were not fully utilized (it is the model evaluation that is multithreaded; the matrix solving is done in a single thread still), so if you have 4 CPUs you might as well ask for 4, and you can be using the unused CPU resource for something else at the same time (it seemed to be using around 2 CPUs-worth on average).

In my case, I got something like 30% speedup, but the circuit was not that large. This was with bsim3v3, but I'd expect the improvement could be more with bsim4 models (since they are more complex).

Note this is based on tests I did 18 months ago, so YMMV

Andrew.