Drift-Tolerant Implementation of a Neural Network on a PCM-Based Analog In-Memory Computing Unit for Motor Control Applications

Neural Networks (NN) are involved in a plethora of applications, such as gesture recognition and data labeling. In this scenario, Analog In-Memory Computing (AIMC) is a promising paradigm to deal with the prohibitive energy constraints and performance issues related to the NN implementation. As a proof of concept, this paper proposes a realization of a simple NN for a motor control driver through an AIMC unit based on an embedded Phase-Change Memory (ePCM), both designed in a 90-nm CMOS STMicroelectronics technology. The accuracy of the NN inference is characterized in different scenarios to evaluate the influence of cells conductance drift in a real application. In particular, classification accuracy in both uncompensated and compensated cases is experimentally evaluated at several time intervals and considering different cells conductance levels. Results highlight that the proposed compensation method allow to obtain an accuracy above the 96 % after 5 days from cells programming even with lower levels of conductance.