Mismatch cost of computing: from circuits to algorithms

25 Nov 2024  ·  Abhishek Yadav, Francesco Caravelli, David Wolpert ·

Stochastic thermodynamics extends equilibrium statistical physics to systems arbitrarily far from thermal equilibrium, with arbitrarily many quickly evolving degrees of freedom. These features make it the correct framework for analyzing the thermodynamics of real-world computers. Specifically, stochastic thermodynamics has been used to prove that the "mismatch cost" of a dynamic process is a lower bound on the energy dissipation, namely the entropy production (EP), of any physical system that implements that process. In particular, the mismatch cost for any periodic process - like every modern digital device - is strictly positive. Here we show that mismatch cost may be significant on the macroscopic scale, not just on the nanoscale (in contrast to many other stochastic thermodynamics bounds). We also show that the mismatch cost of systems that are coarse-grained in time and/or space still provides lower bounds to the microscopic entropy production of running such systems. We then argue that traditional computer science measures of algorithmic efficiency, focused on the resource costs of "space" and "time" complexity, should include a third cost - thermodynamic cost - and that mismatch cost is well-suited to bounding such a cost for arbitrary computational machines. Accordingly, we derive the mismatch cost for running an arbitrary Boolean circuit and for running an arbitrary Random Access Stored Program machine (a simplified model of a microprocessor).

PDF Abstract