Towards More Precision in Approximate Computing

Imperfect Moore’s law scaling has led to recent trends that consider systems that generate incorrect outputs. The emergent field of approximate computing considers deliberate, but controlled, relaxation of correctness for better performance, energy or reliability. While the fast growing popularity of the field indicates the general acceptance, even eagerness, for the idea, there are many hurdles to the practical application of approximate computing in today’s systems that are preventing it from realizing its full potential. Many of these hurdles will naturally be surmounted by the evolution of the field, but it is imperative that a strong fundamental base be established upon which future research directions can be built. In this paper we highlight some of the challenges faced by approximate computing, namely, standardized benchmarks, output error evaluation and reducing programmer burden through automation.