University of Michigan Boosts Edge AI Efficiency by Merging State-Space Models with In-Memory Computing

Researchers at the University of Michigan College of Engineering have developed a hardware-software co-design approach that directly maps state-space models onto in-memory computing architectures, significantly improving energy efficiency and reducing processing latency for artificial intelligence on edge devices. This method enables real-time processing of continuous data streams—such as video or sensor data—allowing high-performance AI to run locally on devices like smartphones, hearing aids, or autonomous vehicle cameras. The findings were published in *Nature Communications*. The study highlights that computations inherent in state-space models can leverage the physical characteristics of in-memory computing systems for efficient execution, overcoming limitations these hardware platforms face when running convolutional neural networks and Transformer models.