Approximating with Input Level Granularity

Approximate computing is a technique for bridging the growing imbalance between computational power and computational needs by trading small amounts of accuracy for large amounts of performance or energy. Current approximate computing techniques are configured to choose how to approximate based either on training inputs that are representative of the “real” input or on occasional runtime checks that compare the exact results to those of the approximation and adjust accordingly. We argue that because these approaches are based on worst or average case behavior, they cannot make the most of each input and thus they are bound to either cause excessive error or leave performance on the table. We introduce input level granularity, an approach that may make it possible to achieve good performance and acceptable accuracy on every input.