Evolutionary Design of Approximate Digital Circuits

Approximate computing exploits the fact that there are many error-resilient applications (such as image processing, deep learning, and data mining) in which the quality of the result can be traded for energy. An open problem is how to effectively approximate hardware and software, i.e., simplifying or modifying digital circuits and programs so that the resulting application error is acceptable for the end-user and the energy requirements are satisfied for challenging deployments on resource-constrained devices. We will present a design automation method capable of creating high-quality implementations of approximate digital circuits.

The method is based on multi-objective genetic programming. It employs various techniques such as relaxed equivalence checking, optimized search strategies, and genetic operators to eliminate scalability problems typical for design automation approaches. The method was used to design an open-source library of approximate adders and multipliers (EvoApproxLib) and other resource-aware computational primitives such as image filters, hash functions, and convolutional neural networks.