Low power and high performance requirements have increased focus on. Approximate Computing which uses designs that approximate the functionality of a precise design while still achieving acceptable quality of results and consuming lower nergy than the precise designs. To perform design automation for approximate designs, modern CAD tools should have the ability to quickly estimate the output quality of designs that include approximate design modules. Previous research on output quality estimation for approximate designs has focused on using an interval based approach which introduces quantization error, or lookup table-based techniques, which mainly emphasize on output quality estimation for approximate combinational circuits and have large overheads for storing the lookup tables for different error metrics. Other works use an unrolling based approach to estimate the output quality which requires large characterization time. In this work, I propose a methodology to estimate the output quality of approximate sequential circuits based on deriving analytical expressions for predicting approximation errors from statistical data gathered from performing limited characterization of the approximate circuits. I show that limited characterization is sufficient to accurately characterize approximation errors since in many cases, the error behavior of approximate circuits follows a pattern. As such, I show that we can achieve high accuracy of prediction for average approximation error, even with this limited characterization. I also demonstrate that the methodology is more scalable and reduces characterization time by 90% on average compared to previous approaches.
University of Minnesota Master of Science thesis. September 2014. Major: Electrical Engineering. Advisor: John A. Sartori. 1 computer file (PDF); vi, 28 pages.
Kapare, Amrut Shivshankar.
A statistical approach to error prediction in approximate sequential circuits.
Retrieved from the University of Minnesota Digital Conservancy,
Content distributed via the University of Minnesota's Digital Conservancy may be subject to additional license and use restrictions applied by the depositor.