A MEMS-BASED ANALOG COMPUTER FOR EDGE AI COMPUTING
David Lin, Johan Reimann, Dorin Calbaza, Robert MacDonald, Zhihui Yang, Abdallah Al, Mohammad Megdadi, Fadi Alsaleem · 2024
Artificial Intelligence computing at the edge is often limited by size, weight and power.Existing approaches, based on digital computers, are inefficient due to the analog-to-digital conversion burdens and other processing bottlenecks inherently present in the von Neumann architecture.To address these challenges, a novel MEMS-based analog computer was developed and shown to achieve 300x reduction in power and 100x improvement in speed compared to a typical digital computer solving the same machine learning signal de-noising problem.The MEMS analog computer is enabled by two key innovations: MEMS continuous-time recurrent neural network and in-situ supervised training.