Development of High-Efficiency AI Implementation Semiconductor Mimicking the Human Brain

Low Power, High Performance, and Commercialized Silicon Enable High Practicality

[Asia Economy Reporter Kim Bong-su] Korean scientists have developed a high-performance semiconductor that mimics the human brain to implement artificial intelligence (AI) with very low power consumption.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 5th that a joint research team led by Professors Choi Yang-gyu and Choi Sung-yul from the Department of Electrical Engineering developed a highly integrated neuromorphic semiconductor that mimics the human brain.

The research team focused on the fact that the human brain performs very complex functions while consuming only about 20 watts (W) of power. They implemented AI functions in hardware by mimicking the human brain. The neuromorphic hardware developed in this way can perform AI functions at ultra-low power, unlike the conventional von Neumann architecture. It is composed of neurons and synapses mimicking the human brain using a single transistor and was fabricated using a commercial silicon standard process, increasing its potential for practical application.

To implement neuromorphic hardware, neurons that generate spikes when certain signals are integrated, and synapses that remember the connectivity between two neurons, just like biological brains, are required. However, neurons and synapses based on digital or analog circuits occupy a large area, limiting integration density. Considering that the human brain consists of about 100 billion (10^11) neurons and 100 trillion (10^14) synapses, improving integration density is necessary for use in actual mobile and Internet of Things (IoT) devices.

Accordingly, various neurons and synapses based on different materials and structures have been proposed, but most cannot be fabricated using standard silicon fine process technology, making commercialization difficult and mass production challenging.

To solve this problem, the research team mimicked the operation of biological neurons and synapses using a single transistor that can be fabricated with widely used standard silicon fine process technology, and simultaneously integrated them on the same wafer (8-inch) to produce a neuromorphic semiconductor.

The fabricated neuromorphic transistor has the same structure as transistors currently mass-produced for memory and system semiconductors. Experiments demonstrated that the transistor can perform memory functions, logic operations, and new neuromorphic operations. By applying a new operating principle to existing mass-produced transistors, they created neuromorphic transistors with the same structure but completely different functions. The neuromorphic transistor was proven for the first time in the world to be implemented in a Janus structure, performing both neuron and synapse functions simultaneously, like a coin with heads and tails at the same time.

The research team’s technology replaced neurons, which were previously composed of complex digital and analog circuits, with a single transistor, dramatically increasing integration density. By integrating synapses with the same structure, process simplification reduced costs. While the planar area required for a conventional neuron circuit is 21,000 units, the newly developed neuromorphic transistor requires less than 6 units, increasing integration density by about 3,500 times.

Based on the fabricated neuromorphic semiconductor, the research team partially mimicked brain functions such as amplification gain control and concurrency judgment, and confirmed the ability to recognize character images and face images.

PhD candidate Han Jun-gyu said, "We demonstrated that neuron and synapse operations are possible using a single transistor based on complementary metal-oxide-semiconductor (CMOS) technology. By simultaneously integrating neurons, synapses, and additional signal processing circuits on the same wafer using commercial CMOS processes, we improved the integration density of neuromorphic semiconductors. This could accelerate the commercialization of neuromorphic hardware."

This research was published in the August online edition of the international journal Science Advances.

This content was produced with the assistance of AI translation services.