Algorithm Architecture Co-Design for Analog In-Memory Computing

Large Language Models (LLMs) have gained significant traction in real-world AI applications. However, LLMs demand large weight capacity, efficient computing, and high-throughput data communication. We will discuss how non-volatile memory, analog mixed-signal design, system architecture, and workloads impact efficiency and performance of Analog In-Memory Computing. Through circuit simulations and hardware-aware training, we demonstrate near-software accuracies in both simulation and hardware.