The most compact and energy-efficient transistor for AI has been created.
Chinese scientists have developed the world's most compact ferroelectric transistor with a 1 nm gate, which reduces energy consumption during data transmission in artificial intelligence systems. The new device operates at low voltage and is compatible with industrial technologies, paving the way for more energy-efficient chips for AI and wearable devices.
Ingenium
A new approach to designing a "nano-gate" makes it possible to bridge the voltage gap between memory and logic elements, which could significantly reduce energy consumption during data transfer in artificial intelligence systems.
A research team from Peking University and the Chinese Academy of Sciences has created the world's most compact ferroelectric transistor, featuring a gate length of just 1 nanometer. This device, detailed in the journal Science Advances, operates at a voltage of 0.6 volts, overcoming one of the main energy limitations in the semiconductor industry.
Modern logic microcircuits typically run at around 0.7 volts, while non-volatile memory such as NAND flash requires 5 volts or more for writing. Even previous ferroelectric field-effect transistors (FeFETs) needed over 1.5 volts. This mismatch in operating voltages necessitates additional voltage-boosting circuits, which increase energy consumption and take up valuable chip space. In typical AI chips, data transfer accounts for 60 to 90% of total energy use, while computation itself consumes much less.
To address this issue, the team used metallic single-walled carbon nanotubes as gate electrodes. This design acts as a nano-tip, concentrating the electric field and strengthening the connection between the ferroelectric layer and the transistor channel.
The enhanced field allows polarization switching at just 0.6 volts—lower than standard logic voltages—while the device remains resistant to short-channel effects.
Ferroelectric transistors based on molybdenum disulfide (MoS2) demonstrate excellent memory characteristics: the on/off current ratio reaches 2 million, and programming speed is 1.6 nanoseconds. Thanks to voltage compatibility between memory and logic, there is no need for extra charge-pumping circuits, enabling rapid data exchange.
The core operating principle of this device can be applied to various ferroelectric materials and is compatible with standard industrial manufacturing processes. This development could play a key role in creating energy-efficient solutions for large AI models, edge computing, and wearable devices, where energy savings are crucial.