Quality-Scalable Approximate LMS Filter

Approximate Computing allows improving circuits performances by accepting inaccuracies in the calculations. Adaptive Least Mean Squares (LMS) filters can benefit from Approximate Computing, since they are inherently inexact and power-hungry. In this paper we propose a Quality-Scalable approximate LMS filter, in which it is possible to change the approximation level at runtime, by acting on an external quality knob. This allows to adapt the filter performance according to the target application and the input data. The proposed approach introduces approximations at algorithmic level. The filter is able to enter automatically in a low-power approximated-mode, by freezing the update of some of its coefficients. A simple auxiliary circuit monitors the error and the quality knob and drives the filter in the approximated-mode of operation when convergence has been reached. The proposed approximate LMS filter is implemented in TSMC 40nm CMOS technology. The implementation results show a power improvement in the range 5% to 32%, as a function of the desired quality loss.