Reconfigurable Computing

Chapter 23. Precision Analysis for Fixed-point Computation

George A. ConstantinidesDepartment of Electrical and Electronic EngineeringImperial College, London

Many values in a computation are naturally represented by integers, which have very efficient hardware implementations; basic operations are relatively cheap, and they map well to an FPGA’s underlying hardware. However, some computations naturally result in fractional values, that is, numbers where part or all of the value are less than 1—for example, 0.25, 3.25, and π—or that are so large that representation as integers is too costly—for example, 10¹²⁰. Handling these values is a significant concern because the hardware necessary to compute on scaled values can be significant in speed, power ...