Heterogeneously Integrated Intelligent System for Learning at the Edge

Edge learning deploys training directly on IoT devices to enable real-time, on-device adaptation. The edge learning reduces data movement and processing overhead, but puts high demand on computing capability of edge devices. In this paper, we discuss a unified heterogeneous integration architecture, building on our recent advancements in 3D memristors, ECRAM arrays, and 3D hetero-integration techniques. With 3D memristor arrays for convolutional preprocessing and ECRAM arrays for efficient in-situ learning, the system achieves high-throughput, low-power processing suitable for real-time applications.