Neuromorphic systems are hardware architectures that emulate the structure of the
human neural network to perform parallel computation. Based on the integration of memory
and processor units, such systems can mitigate the memory-processor bottleneck, ...
Neuromorphic systems are hardware architectures that emulate the structure of the
human neural network to perform parallel computation. Based on the integration of memory
and processor units, such systems can mitigate the memory-processor bottleneck, thereby
improving computational speed and reducing power consumption.
In these systems, synaptic devices play a key role in modulating the connection strength
between neurons, where high integration density and low-power operation are essential
requirements. Accordingly, extensive research has been conducted on various non-volatile
memory-based synaptic devices, including resistive random-access memory (RRAM),
ferroelectric field-effect transistors (FeFETs), and silicon-oxide-nitride-oxide-silicon
(SONOS) structures.
Among these devices, the SONOS device achieves stable charge retention and excellent
linear conductance modulation through charge trapping and detrapping via strong electricfield-induced Fowler-Nordheim (FN) tunneling. However, its high-field-dependent
operation leads to high operating voltages and slow switching speeds.
In contrast, research on volatile synaptic devices capable of short-term memory and highspeed learning has also been actively conducted. A representative example is the onetransistor dynamic random-access memory (1T-DRAM) structure, which modulates the
channel conductance by charge accumulation in the floating-body region, thereby achieving
fast operation and high energy efficiency. Nevertheless, due to its volatile nature, the
retention time is short, making long-term weight storage difficult.
In this study, a single-transistor synaptic device that combines both mechanisms is
proposed. The proposed hybrid single-transistor synaptic device performs high-speed
learning by modulating the channel conductance through the floating-body effect induced
by gate-induced drain leakage (GIDL) current, while achieving long-term memory by
storing charge in the dielectric layer via FN tunneling.
Using technology computer-aided design (TCAD) simulations, the electrical
performance of the proposed device was thoroughly investigated in terms of threshold
voltage variation, conductance, and sensing margin under both volatile and non-volatile
operating modes. In addition, the obtained device-level data were incorporated into the
NeuroSim framework to extend the analysis to the circuit and system levels, and the
neuromorphic recognition capability was evaluated using the MNIST (Modified National
Institute of Standards and Technology) handwritten digit dataset