Single-device in-sensor computing for multi-channel multiply-accumulate operations

Abstract

Recent advances in in-sensor computing demonstrate the potential of integrating sensing and computation at the perception front end; however, many existing approaches rely on customized devices, facing scalability, uniformity, and power challenges. Here, we present a cross-platform in-sensor computing strategy that embeds multiply–accumulate operations directly into the physical sensing process. We identify a mathematical isomorphism between carry propagation in computation and sensor’s exponential decay response, establishing a general LinExp-τ principle that enables diverse sensing devices to function as efficient computational units. The approach is experimentally validated across different devices, including semiconductor photodetectors, oxide transistors, and conductive polymer-based neural probes. A high-speed photodetector implementation achieves gigascale operation rates with high area efficiency and zero static power consumption during core operations. Beyond device-level performance, the architecture enables real-time visual preprocessing and in-sensor physiological signal solving, including event-driven functional near-infrared spectroscopy. These results provide a scalable, energy-efficient framework for sensing-as-computing systems. In-sensor computing integrates sensing and computation but faces challenges in the uniformity and signal integration of multiple devices. Wang et al. propose a cross-platform strategy that extracts a unified mathematical feature from diverse physical responses, enabling various sensors to function as efficient computational units.