Block-LMS and RLS adaptive filters using in-memory architectures

Deep in-memory architecture (DIMA) provides considerable latency and throughput improvements 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). Past work [1] used adaptive filtering as a means to analyse the impact of nonlinear read operation on gradient descent methods implemented on DIMA. Gradient descent techniques are typically implemented in block fashion in DIMA since write operations have high overhead. This work uses block least mean square (BLMS) adaptive filters as an example to study the impact of nonlinearity and variations encountered in DIMA. The insights gained from BLMS adaptive filters are used to realize recursive least square (RLS) adaptive filters. The steady state mean square error (MSE) of both filters remains unaffected while convergence rate and stability issues can arise when these filters are implemented on DIMA. The insights are useful in context of learning systems employing gradient descent techniques.