University of Virginia
Imprecise hardware challenges the traditional notion that correctness is an immutable priority, by systematically trading off efficacy (precision) against efficiency (power, area, performance, and cost). Evaluating the impact of such tradeoffs on output quality using, e.g., Monte Carlo simulations is a time-consuming and non-deterministic process. This paper presents two analytic modeling techniques for evaluating error properties and output quality in imprecise arithmetic circuits, based on interval arithmetic and affine arithmetic. Experiments show that these techniques offer significant speedups over previous methods, as well as promising estimation accuracy.