One of the big problems with the development of AI models is that the massive amount of computation requires power.  Removing the heat generated by consuming all that power requires yet more power. A potential solution uses an information-theoretic concept called reversible computing. 

The idea is simple, but the implementation is hard.  In reversible computing, the energy that is used to perform a logic operation can be recovered by reversing that operation. For a light switch, we use energy each time we flip the switch on or off.  But if we spring load the switch with a latch to turn it on, then the potential energy in the spring can be recovered when the switch is turned off. In reversible computing, the energy used in computations (AND and NOT gates) are continually recovered by reversing the computation, using specially-designed gates where the computation can be logically reversed.  The trick is in being able to use that recovered energy for the next set of computations. The answer is to temporarily store the energy in a resonator formed from inductance/capacitance (LC) circuits, so that energy oscillates from the logic gates to the resonator and then back again, to perform the next set of computations in the program. The computation may take twice as long because the steps have to be reversed, but the amount of energy required to perform the computation can be greatly reduced.  

A recent article in IEEE Spectrum provides more details. The idea has reached current attention because a start-up, Vaire.co, founded by Rodolfo Rosini and Dr, Hannah Earley, has assembled a team of experts in both the theory and microelectronics design, and has begun developing prototype chips to prove out the technology. Their new chip, the “Ice River” chip, saves 30% in energy consumption, as recently reported in Science News.  Theoretically, much more is possible. 

Lower heat dissipation should allow for greater density and faster computing cycles, even with the overhead of reversing computations to recover energy.  It might solve the problem of requiring new energy sources for AI computing, and make practical computing centers in space, where heat dissipation is problematic.