Abstract The rapid growth of artificial intelligence and the Internet of Things calls for compact hardware platforms that integrate sensing, computing, and nonlinear processing within a unified architecture. However, most existing neuromorphic systems implement only partial functionalities and rely on heterogeneous device integration, limiting scalability and efficiency. Here, we show a high-speed, reconfigurable multi-modal split-floating-gate memory that monolithically integrates in-sensor computing, in-memory computing, and multiple nonlinear activation functions within a single device structure. By programming charges in spatially separated floating gates, the device enables non-volatile analog control of photoresponsivity and conductance, as well as electrically reconfigurable rectification to emulate ReLU and Sigmoid activations. We further demonstrate a fully hardware-implemented sensor–processor system based on the multi-modal split-floating-gate memory arrays that performs complete unsupervised and supervised learning tasks. This work establishes a compact, energy-efficient, and reconfigurable hardware foundation for scalable intelligent systems beyond conventional silicon architectures.
Zhang et al. (Wed,) studied this question.