A South Korean research team has developed a new AI semiconductor device that mimics the learning principles of the human brain [1].

This development addresses a critical bottleneck in AI hardware. Current chips often face a trade-off where increasing electrical current to improve performance reduces storage efficiency, or vice versa [1]. By overcoming this limitation, the new device allows for the processing of larger datasets while consuming significantly less power [1].

The research team, which includes graduate student Shin Hee-sung of Korea University, focused on emulating synaptic plasticity [1]. Synaptic plasticity is the biological process by which the brain strengthens or weakens connections between neurons based on activity. By replicating this mechanism, the semiconductor can operate with higher efficiency than traditional linear architectures [1].

To achieve this high-performance, low-power operation, the team combined two specific materials: ferroelectric hafnium-zirconium oxide (HZO), and indium gallium zinc oxide (IGZO) [1], [2]. The integration of these materials allows the device to maintain stable data storage while facilitating the rapid transitions required for AI processing [2].

Industry experts note that the competition to develop efficient hardware is intensifying as AI models grow in complexity. An anchor for YTN said, "As AI technology advances, the competition to develop semiconductors that can process more data with less power is becoming more intense" [1].

The team's approach focuses on the physical layer of the chip to reduce the energy overhead typically associated with moving data between memory and processors. This neuromorphic approach aims to bring the computing process closer to the way biological systems function, potentially reducing the heat and energy waste found in current data centers [1].

“The new device allows for the processing of larger datasets while consuming significantly less power.”

The shift toward neuromorphic computing—hardware that mimics the brain's structure—is a response to the energy crisis facing large-scale AI. By utilizing HZO and IGZO materials to bypass the traditional power-performance trade-off, this research suggests a path toward 'edge AI' where complex processing can occur on small devices without relying on massive, energy-hungry server farms.