Evolutionary design of approximate circuits
Lukas Sekanina
Faculty of Information Technology, Brno University of Technology
DEIB - NECSTLab Meeting Room (Building 20, basement floor)
October 28th, 2019
2.30 pm
Contacts:
Marco Santambrogio
Research line:
System architectures
Abstract
Approximate computing exploits the fact that there are many error-resilient applications (such as image recognition, video processing and data mining) in which the quality of service can be traded for performance or power consumption. An open problem is how to effectively approximate hardware and software, i.e. how to simplify or modify circuits and programs in such a way that the resulting application error is acceptable for the end user and significant improvements in terms of performance and/or energy are obtained. This talk introduces a genetic programming based method developed for an automated design of approximate circuits. We will present a library (called EvoApprox) of approximate adders and multipliers, containing more than 20,000 circuit instances, automatically generated by the method. Some of these approximate multipliers were evaluated in an image classification system based on convolutional neural network, in which power consumption of arithmetic operations is traded for a negligible increase in the classification error. The proposed method was also used to automatically generate energy-efficient image filters. In order to exactly determine the quality (expressed as the average error, the worst-case error etc.) of candidate solutions, symbolic methods based on BDD analysis and SAT problem solving are employed. This topic will also be briefly introduced in the talk.The NECSTLab is a DEIB laboratory, with different research lines on advanced topics in computing systems: from architectural characteristics, to hardware-software codesign methodologies, to security and dependability issues of complex system architectures.