Přístupnostní navigace
E-application
Search Search Close
Publication detail
SEKANINA, L. MRÁZEK, V. VAŠÍČEK, Z.
Original Title
Design Space Exploration for Approximate Implementations of Arithmetic Data Path Primitives
Type
conference paper
Language
English
Original Abstract
While a detailed analysis of various approximation strategies for elementary arithmetic circuits (such as adders and multipliers) is widely covered in the literature, much less is known about approximate arithmetic data path primitives (such as Dot Product (DP) to name one example) because it is difficult to exactly analyze their error and other parameters. We provide a detailed analysis of approximation options if a two-dimensional DP circuit is implemented using pre-characterized approximate arithmetic circuits available in existing libraries of approximate circuits. We propose a method capable to cheaply estimate properties of candidate approximate implementations of DP circuit. By a careful selection of approximate components, hundreds of approximate DP circuits showing excellent tradeoffs between key design objectives were obtained.
Keywords
approximate circuit, dot product, Pareto front, parameter estimation
Authors
SEKANINA, L.; MRÁZEK, V.; VAŠÍČEK, Z.
Released
7. 12. 2018
Publisher
IEEE Circuits and Systems Society
Location
Bordeaux
ISBN
978-1-5386-9562-3
Book
25th IEEE International Conference on Electronics Circuits and Systems (ICECS)
Pages from
377
Pages to
380
Pages count
4
URL
https://www.fit.vut.cz/research/publication/11765/
BibTex
@inproceedings{BUT155069, author="Lukáš {Sekanina} and Vojtěch {Mrázek} and Zdeněk {Vašíček}", title="Design Space Exploration for Approximate Implementations of Arithmetic Data Path Primitives", booktitle="25th IEEE International Conference on Electronics Circuits and Systems (ICECS)", year="2018", pages="377--380", publisher="IEEE Circuits and Systems Society", address="Bordeaux", doi="10.1109/icecs.2018.8618048", isbn="978-1-5386-9562-3", url="https://www.fit.vut.cz/research/publication/11765/" }
Documents
ICECS 2018 paper.pdf