Embedded Algorithmic Noise-Tolerance for Signal Processing and Machine Learning Systems via Data Path Decomposition

Low overhead error-resiliency techniques such as algorithmic noise-tolerance (ANT) have been shown to be particularly effective for signal processing and machine learning kernels. However, the overhead of conventional ANT can be as large as 30% due to the use of explicit estimator block. To overcome this overhead, embedded ANT (E-ANT) is proposed [S. Zhang and N. Shanbhag, Embedded error compensation for energy efficient DSP systems, in Proc. GlobalSIP, Dec. 2014, pp. 30-34], where the estimator is embedded in the main computation block via data path decomposition (DPD). E-ANT reduces the logic overhead to be below 8% as compared with the 20%-30% associated with conventional reduced precision replica (RPR) ANT system while maintaining the same error compensation functionality. DPD was first proposed in our original paper [Zhang and Shanbhag, 2014] where its benefits were studied in the case of a simple multiply-accumulator (MAC) kernel. This paper builds upon [Zhang and Shanbhag, 2014] by 1) providing conditions for the existence of DPD, 2) demonstrating DPD for a variety of commonly employed kernels in signal processing and machine learning applications, and 3) evaluating the robustness improvement and energy savings of DPD at the system level for an SVM-based EEG seizure classification application. Simulation results in a commercial 45 nm CMOS process show that E-ANT can compensate for error rates up to 0.38 for errors in FE only, and 0.17 for errors in FE and CE, while maintain a true positive rate p(tp) > 0.9 and a false positive rate p(fp) <= 0.01. This represents a greater than 3-orders-of-magnitude improvement in error tolerance over the conventional system. This error tolerance is employed to reduce energy via the use of voltage overscaling (VOS). E-ANT is able to achieve 51% energy savings when errors are in FE only, and up to 43% savings when errors are in both FE and CE.