Fused Multiply-Add for Variable Precision Floating-Point

Abstract

In this work, we address the design of a Fused Multiply-Add (FMA) in Tunable Floating-Point (TFP). TFP is a floating-point variable precision format in which a given precision for significand and exponent can be chosen for a single operation. The objective is to increase the power efficiency of the computation by tuning the precision of algorithms that can tolerate some error. The performance of the FMA is compared to that of separate multiply and add units on computation kernels used in several applications.