Linear compensation of drift and mitigation of drift variations in PCM arrays for highly accurate, energy efficient in-memory computing

Phase change memory (PCM) offers high density and low current operation for energy efficient in-memory computing (IMC). However, conductance drift is still a major challenge affecting the accuracy of IMC-based edge artificial intelligence (AI) accelerators. This work presents a new weight mapping technique to linearly compensate drift, thus ensuring the linearity of matrix-vector multiplication (MVM). We show a new integrated circuit for MVM capable of current subtraction in the analog domain. High inference accuracy is demonstrated for mapped deep neural networks (DNNs) even after extensive annealing, thanks to differential mapping and optimization of read voltage V<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">read</inf> to mitigate drift-induced variations. These results support drift-compensated PCM as a robust technology for IMC-based edge AI inference.