What is the minimum power requirement to preform a computation? The answer is none, if the computation is reversible. So computing f(a, b) ≝ a + b consumes energy because the loss of information must be dissipated as heat. However, computing g(a, b) ≝ (a + b, a − b) could be done without consuming any energy because the inputs can be recomputed from the output. And furthermore all computation can be turned into reversable computation by making a copy of the input in the output.
Of course this is a bit strange because the result of g contains more information that the result of f. There is something strange going on between information, entropy, computation, and energy that I don’t really understand.
The real energy cost seems to be in initializing the input to the values you want. Setting registers is not a reversible process.