Processing circuitry designed to allow imprecision can be up to 15 times more power efficient than conventional circuitry for some tasks. The technique is particularly suitable for audio and graphics subsystems, and researchers say that such circuits could start showing up in produtss such as hearing aids and tablet computers in 2013. The basic concept, which has ties to fuzzy logic and probabilistic processing, is simple: allow hardware for operations such as multiplication and addition to sometimes generate incorrect results, but manage the probability of error accumulation.

A prototype processor embodying this concept has been built by a team of researchers from Rice University (Houston), Nanyang Technological University (Singapore), the Centre for Electronics and Microtechnology (Switzerland) and the University of California (Berkeley). The team used a number of techniques to deviate from conventional full-precision circuits. These include pruning rarely used portions of digital circuits and  confined voltage scaling.

In initial simulations the team found that pruned versions of conventional circuits could run twice as fast, be half as large and use half the power of the original circuits. In the latest research the team implemented their ideas in a prototype silicon chip. They showed that pruning could cut energy use by a factor of 3.5 with chips that deviate from the correct value by 0.25% on average. When size and speed gains were factored in, the chips were 7.5 times more efficient than regular chips. Chips with a larger deviation of about 8% were up to 15 times more efficient.

Initial applications for pruning are likely to be in application-specific processors for hearing aids, cameras and other electronic devices. Inexact hardware is also being considered for IS Aid’s I-slate educational tablet designed for Indian classrooms. Pruned chips are expected to cut power requirements in half and allow the I-slate to run from small solar panels similar to those used on handheld calculators.


Image: Rice University