Evaluation of the Functional Impact of Approximate Arithmetic Circuits on Two Application Examples

Approximate arithmetic units emerge as an attractive alternative solution for high-performance, low-power, intensive computing applications. The error acceptability and resilience of such an inexact solution is very dependent on the application field. In this chapter, a set of well-known approximate adders (TrA, SOA, LOA, GeAr) and multipliers (UDM, BAM, AMB, LM) are presented, and their impact is evaluated in the context of two application examples: FIR digital filters and convolutional neural networks for object detection (YOLO). For both applications, it is analyzed how their functional performances are affected by the usage of approximate arithmetic units. In addition to this pure functional analysis, for the convolutional neural networks, the impact on the silicon area, delay, and power dissipation is also discussed.