Precision-scalable techniques constitute an efficient solution to power consumption issues thanks to the possibility to adapt arithmetic components precision to required system-level accuracy with the aim to dynamically optimize power consumption. In this paper we propose a precision-scalable approach for the implementation of a Least Mean Square (LMS) filter. Novel solution exploits variable rounding multiplications in the learning section of the LMS filter allowing to dynamically reduce the switching activity of multipliers partial products with a minimal impact on error regime performance. Results, obtained after a Place & Route in TSMC 28 nm CMOS technology, reveal a regime precision comparable to a standard LMS implementation and a power consumption improvement up to 27%.
展开▼
