Non Volatile Memory

Non-volatile memory (NVM) is being actively researched for its potential to revolutionize computing, particularly in accelerating deep neural networks (DNNs) through in-memory computing (IMC). Current research focuses on mitigating NVM's inherent imperfections, such as variations and noise, through algorithmic and architectural co-design, including techniques like write-verify optimization and negative feedback training, often applied to transformer and convolutional neural network models. These efforts aim to improve the accuracy, energy efficiency, and reliability of NVM-based DNN accelerators for applications ranging from edge AI to safety-critical systems, ultimately impacting both hardware design and machine learning algorithms.