Memory error compensation techniques for JPEG2000

摘要

This paper presents novel algorithm-specific techniques to mitigate the effects of failures in SRAM memory caused by voltage scaling and random dopant fluctuations in scaled technologies. We focus on JPEG2000 as a case study. We propose three techniques that exploit the fact that the high frequency subband outputs of the discrete wavelet transform (DWT) have small dynamic range and so errors in the most significant bits can be identified and corrected easily. These schemes do not require any additional memory and have low circuit overhead. We also study several error control coding schemes that are effective in combating errors when the memory failure rates are low. We compare the PSNR versus compression rate performance of the proposed schemes for different memory failure rates. Simulation results show that for high bit error rates (10⁻²), the error control coding techniques are not effective and that the algorithm-specific techniques can improve the PSNR quality of up to 10dB higher compared to that of the no-correction case.