Reversibility for efficient computing

Abstract

Today's computers are based on irreversible logic devices, which have been known to be fundamentally energy-inefficient for several decades. Recently, alternative reversible logic technologies have improved rapidly, and are now becoming practical. In traditional models of computation, pure reversibility seems to decrease overall computational efficiency; I provide a proof to this effect. However, traditional models ignore important physical constraints on information processing. This thesis gives the first analysis demonstrating that in a realistic model of computation that accounts for thermodynamic issues, as well as other physical constraints, the judicious use of reversible computing can strictly increase asymptotic computational efficiency, as machine sizes increase. I project real benefits for supercomputing at a large (but achievable) scale in the fairly near term. And with proposed future computing technologies, I show that reversibility will benefit computing at all scales. Next, the thesis demonstrates that reversible computing techniques do not make computer design much more difficult. I describe how to design asymptotically efficient processors using an "adiabatic" reversible electronic logic technology that can be built with today's microprocessor fabrication processes. I describe a simple universal reversible parallel processor chip that our group recently fabricated, and a reversible instruction set for a more traditional RISC-style uniprocessor. Finally, I describe techniques for programming reversible computers. I present a high-level language and a compiler suitable for coding efficient reversible algorithms, and I describe a variety of example algorithms, including efficient reversible sorting, searching, arithmetic, matrix, and graph algorithms. As an example application, I present a linear-time, constant-space reversible program for simulating the Schrodinger wave equation of quantum mechanics.