Development of Ultra-Low Power Semiconductor Devices Using 'Seukemion'

[Asia Economy Reporter Junho Hwang] An artificial intelligence semiconductor (AI) device that electrically controls skyrmions, spin structures at the nanoscale, and utilizes them like human neural networks has been developed. While it is significant that a method to control skyrmions and apply it to semiconductor devices was developed, the research is also meaningful because the device can greatly reduce power consumption compared to existing semiconductors.

On the 29th, the Korea Institute of Science and Technology (KIST) announced the research results of a joint research team including Dr. Kyungmi Song from the Next-Generation Semiconductor Research Center, Dr. Hyunsoo Joo, Director Junyeon Jang, and Dr. Seunghoon Woo from IBM.

The research team developed a method to electrically control skyrmions, spin structures arranged in a vortex shape. This technology can electrically control skyrmions by adjusting their number to change synaptic weights, based on the same principle as neurotransmitters.

Based on this technology, the team fabricated the first synapse device. Using this artificial synapse device, they conducted learning for recognizing handwritten digit patterns (MNIST). After 15,000 pattern learning iterations, they achieved a 90% recognition rate. This recognition rate typically requires hundreds of thousands of repeated learning cycles with existing artificial synapse devices. Power consumption was also reduced to about one-tenth compared to existing devices.

Dr. Kyungmi Song of KIST said, "This is the world's first implementation of an artificial synapse device based on skyrmions, which had only been proposed theoretically before. By controlling synaptic weights according to the number of electrically controlled skyrmions, it closely mimics the human brain's adjustment of synaptic weights through the amount of neurotransmitters."

Dr. Hyunsoo Joo of KIST stated, "The new approach using 'skyrmions' in this research presents a new method for neuro-morphic devices based on next-generation materials or new device platforms, which has significant implications for research in this field."

This research result was published in the world-renowned journal Nature Electronics.