Correcting Sign Calculation Errors in Configurable Approximations

Approximate computing is expected to be a promising technique that achieves both high performance and low power. Addition is an essential operation in many applications and thus approximate adders have been widely studied. Unfortunately, most approximate adders handle only unsigned numbers and thus sign calculation errors occasionally cause destructive results in some applications. This paper targets Carry-Maskable Adder (CMA) and investigates a technique that corrects the errors by exploiting dynamic configurability. Three schemes predicting sign calculation errors are proposed and are evaluated using a handwritten digit recognition application on a deep neural network. It is found that CMA with the best prediction scheme recovers 22.85 percent point of the recognition rate and it still has power reduction of 10.1%.