Unum Arithmetic: Better Math with Clearer Tradeoffs

Abstract

To raise performance beyond Moore’s law scaling, Approximate Computing reduces arithmetic quality to increase operations per second or per joule. It works on only a few applications. The quality-speed tradeoff seems inescapable; however, Unum Arithmetic simultaneously raises arithmetic quality yet reduces the number of bits required. Unums extend IEEE floats (type 1) or provide custom number systems to maximize information per bit (type 2). Unums achieve Approximate Computing cost savings without sacrificing answer quality.

Publication
PMES 2016