| In this dissertation, we study the problem of reliable and energy-efficient digital signal processing (DSP) systems. We present a general framework for deep submicron (DSM) noise/soft error-tolerance and provide its energy-distortion analysis. We introduce a novel algorithmic noise tolerance (ANT) technique referred to as reduced precision redundancy (RPR), which can be easily integrated to any DSP system. Combined with voltage overscaling (VOS), RPR achieves desired performance with considerable energy savings. We extend forward predictor-based ANT and present hybrid ANT and forward-backward predictor ANT as a way to ameliorate the performance loss due to error propagation and burst error. Moreover, by exploiting the statistical properties of VOS error, we propose a direct error estimation and correction technique based on the maximum a posteriori (MAP) principle. We show that zero error operation can be achieved by a MAP-based error correction technique. We also extend the ANT principle and propose an error control technique for the soft error environment referred to as algorithmic soft error-tolerance (ASET). The proposed ANT and ASET approaches provide substantial reduction in energy dissipation while introducing enormous flexibility in the design space. |
