I do not know whether the current ngspice or the other
couple of simulators support veriloga, but a structural
veriloga which includes transition-charge and DC leakage
could simulate efficiently, from the same schematic or
top level netlist.
I'd expect that a good power estimate would have to
assume a particular code and data stream. That would
then be a whole lot of thrash for a real long time. You
want to get a good-enough-for-decision figure of merit,
not 10-bit accuracy at every stinkin' timestep.