Bio-inspired error-tolerant and energy-efficient signal processing

Bio-inspired;Population coding;Energy-efficient;Error-tolerant;Voltage over scaling

Zhang, Gong ; Shanbhag ; Naresh R. ; Jones ; Douglas L.

关键词: Bio-inspired;    Population coding;    Energy-efficient;    Error-tolerant;    Voltage over scaling;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/42135/Gong_Zhang.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

In many applications at the sensory edge, such as security and environmental sensing, reliable sensor nodes must operate for extended time periods on battery supplies. To meet this constraint, energy-efficient systems havebeen developed through different technologies. The primary and the most effective approach has been technology scaling. Another emerging technique is to operate circuits in the subthreshold region as some of the applicationssuch as environmental sensing do not require high throughput. However, both techniques lead to large process, voltage and temperature variationsand therefore jeopardize system reliability.In order to achieve both energy savings and system reliability, we take inspirations from biological methods, such as population-coding, and apply these methods to a canonical problem of non-coherent (unknown phase) frequency estimation. Energy efficiency is achieved using low cost, overlappingband-pass filters rather than conventional non-overlapping band-pass filters.Energy savings are also achieved by operating the hardware at a voltage lower than the nominal voltage (voltage overscaling), which leads to hardwaretiming errors. In the presence of these hardware errors, signal statistics are generated from overlapping band-pass filters with frequency redundancy. Robusttechniques, such as median estimation and algorithmic noise-tolerance, are applied to filter outputs to achieve error-tolerance. Energy/performancetrade-offs are further explored by altering the supply voltage. Simulation results show that the root-mean-squared-error of the bio-inspired method can be reduced by an order of magnitude relative to that of the conventional architecture while achieving an energy consumption reduction of 78% relative to the conventional method which is under hardware-error-free operations atnominal supply voltage.