Adaptive Filtering in In-Memory-Based Architectures

Abstract

Deep in memory architecture (DIMA) has been proposed as a means to improve energy efficiency and latency over conventional digital architectures. DIMA reads multiple bits per bit-line (BL) in each cycle, and performs mixed-signal processing at the periphery of the bit cell array (BCA). While DIMA provides considerable performance benefits, the multi-row read is a non-linear operation. This work studies the impact of non-linearity and variations on stochastic gradient descent (SGD). The analysis is carried out in the context of LMS adaptive filters. The steady state MSE of the filter remains unaffected while convergence rate depends parameters associated with DIMA read operation. The insights are useful in context of learning systems employing gradient descent techniques.