Profile-Based Output Error Compensation for Approximate Arithmetic Circuits
- 1 June 2020
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 67 (12) , 4707-4718
Abstract
Truncation is one of the most commonly used approaches for circuit-level approximate computing. This paper proposes a scheme for error compensation of arithmetic circuits in which a so-called padding is utilized to compensate at the output for the truncated bits of the input operands. Compensation relies on adjusting the output results of an arithmetic circuit; the padding takes a value determined by utilizing statistical information based on profiling an arithmetic circuit to reduce the average signed difference between the inexact and exact values and so the mean square error. An extensive analysis and simulation-based evaluation of error metrics are performed on signed truncated adders, multipliers and dividers; an excellent agreement is found. Additional design metrics such as power consumption and circuit complexity are also assessed. Different applications of approximate arithmetic circuits with the proposed output error compensation scheme are presented. Matrix multiplication and image processing (changing detection) are investigated to show the effectiveness of the scheme proposed in this paper.Keywords
Funding Information
- NSF (1812467)
- National Natural Science Foundation of China (61871216)
- Natural Sciences and Engineering Research Council of Canada (RES0025211)
This publication has 25 references indexed in Scilit:
- Design of Approximate Radix-4 Booth Multipliers for Error-Tolerant ComputingIEEE Transactions on Computers, 2017
- Design-Efficient Approximate Multiplication Circuits Through Partial Product PerforationIEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2016
- On the Design of Approximate Restoring Dividers for Error-Tolerant ApplicationsIEEE Transactions on Computers, 2015
- Improved matrix multiplier design for high‐speed digital signal processing applicationsIET Circuits, Devices & Systems, 2014
- Approximate XOR/XNOR-based adders for inexact computing2013
- Approximate computing: An emerging paradigm for energy-efficient design2013
- New Metrics for the Reliability of Approximate and Probabilistic AddersIEEE Transactions on Computers, 2012
- A new circuit simplification method for error tolerant applications2011
- ERSA: Error Resilient System Architecture for probabilistic applications2010
- Algorithmic transforms for efficient energy scalable computation2000