The increasing demand for artificial intelligence (AI) is driving the expansion of on-device AI in various edge platforms. To overcome the limitations of traditional computing architecture and realize energy-efficient AI at the edge, neuromorphic syst...
The increasing demand for artificial intelligence (AI) is driving the expansion of on-device AI in various edge platforms. To overcome the limitations of traditional computing architecture and realize energy-efficient AI at the edge, neuromorphic systems with integrated memory and processors have gained significant attention. In particular, on-chip training neuromorphic systems, capable of independent learning without server connections, are being actively researched. In such systems, synaptic devices that store weights in the form of analog conductance are required to have characteristics such as low power consumption and high cycling endurance, which are different from requirements for conventional nonvolatile memories like NAND flash. This dissertation proposes a new synaptic device that utilizes lateral charge migration in the Si3N4 charge trap layer, diverging from traditional flash memory devices. The proposed device exhibits faster erase speed and higher reliability compared to other devices and enables low-power selective weight updates in a NOR-type array suitable for on-chip backpropagation. To deepen the understanding of lateral migration-based devices, quantitative measurements and physical analyses were conducted, and the lateral charge migration during programming was shown to be driven by thermally-assisted tunneling emission of electrons under a strong electric field for the first time. Furthermore, the proposed device was successfully integrated with CMOS peripheral circuits, demonstrating the feasibility of a full neuromorphic hardware implementation. High-level simulations validated that on-chip training efficiently compensates for device nonidealities, enabling high system accuracy.