The Community for Technology Leaders
Green Image
Issue No. 09 - September (2011 vol. 60)
ISSN: 0018-9340
pp: 1313-1326
Mudassar M. Nisar , Georgia Institute of Technology, Atlanta
Abhijit Chatterjee , Georgia Institute of Technology, Atlanta
In many DSP applications (image and voice processing), several dBs of SNR loss can be tolerated without noticeable impact on application level performance. For power optimization in such applications, voltage overscaling (VOS) can be used to operate the arithmetic circuitry at or marginally below the critical circuit path delay while incurring tolerable SNR loss due to the resulting periodic errors in computation. In this paper, low cost checksum codes are used for detection and compensation of intermittent errors due to voltage overscaling in linear digital filters. In traditional coding theory, diagnosis of errors is a key problem and incurs significant computation and latency cost. In the proposed approach, low-precision shadow latches are used to identify likely sources of errors due to voltage overscaling to avoid error diagnosis. This allows accurate error compensation with distance-2 checksum codes that are normally good only for error detection but not for correction. Very precise compensation is achieved by distributing the negative of the error value evenly across only the likely erroneous states. This is called guided probabilistic compensation, as compensation is not exact when errors occur simultaneously in more than one state. A feedback controller is used for dynamic voltage overscaling (DVOS) while keeping the error rate in the system within an acceptable range. It is shown that the low cost accurate error compensation allows significant power savings with minimal degradation in system performance (SNR).
Error correcting codes, voltage scaling, guided compensation, low-power filter.

M. M. Nisar and A. Chatterjee, "Guided Probabilistic Checksums for Error Control in Low-Power Digital Filters," in IEEE Transactions on Computers, vol. 60, no. , pp. 1313-1326, 2010.
95 ms
(Ver 3.3 (11022016))