HOME UP PREV NEXT (Typical Numerical Values)

Landauer Limit and Reversible Computation

There are theoretical limits on the energy an irreversible computation requires. Current technology is a long way off these in two resepects:

we use too much energy representing and communicating bits and
we use Vonn Neumann based computation which does not scale well.

Let's consider electrical computers:

If we build a computer using a network of standard components (such as transistors), where the inteconnection pattern expresses our design intent, then the components must be at different spatial locations. The computer will have some physical volume.
If we interconnect our components using electrical wires these will have capacitance, resistance and inductance that stop them behaving like ideal conductors. The smaller the volume, the less wire we need and the better the wires (and hence computer) will work.
If we use transistors that need a swing of about 0.7 volts on their gates to switch them reliably between off and on, then our wires need to move through at least that much potential difference.

The capacitance of the wires turns out to be our main enemy. Given a prescribed minimum voltage swing, the energy used by switching a wire between logic levels can only be made smaller by reducing its area and hence capacitance. Hence smaller computers are better.

Relative chart of dynamic energy use.

»Landauer worked out the miniumum energy use, in theory, for a computer that deletes data as it goes (e.g. erasing the old content of a register when new data is loaded). Computing more efficiently than this requires major low-level design changes so that information is not deleted all time, taking us towards reversible computing.

The traditional approach of wasting the energy of each transition can be countered by returning the charge to the power supply »(eg. Asynchrobatic logic for low-power VLSI design by David John Willingham, Univ Westminster, 2010) and reversible logic (eg. Toffoli logic) can get below the Landauer limit. Some standard algorithms such as encryption and lossless compression are reversible at the large - the trick is mooted to be to code in a way that does not delete intermediate results during the computation. Such techniques may be in wide use within a decade or perhaps two.

The figure above shows ballpark figures for dynamic energy use in today's (2016/2018) mainstream 28 nm silicon technology. We see that contemporary computers are about six orders of magnitude above the Landauer limit in terms of energy efficiency, so there is a lot of improvement still possible before we have to consider reversibility. (Log to base 18 months of one million is roughly 30 years, so 2050 is the reversible computing deadline.)