Limits of AlphaGo That Defeated Lee Sedol, Surpassed by the 'Artificial Brain' [Reading Science]

Rapid Rise of Neuromorphic Semiconductor Technology Mimicking Human Brain Structure
Active Research on 2nd Generation Technology Using New Materials Following 1st Generation Circuit Methods
Will It Become an 'Alternative' to Overcome the Limitations of Existing Von Neumann Computing, the 'Resource-Energy Hog'?

[Asia Economy Reporter Kim Bong-su] Technologies of the 4th Industrial Revolution era, such as artificial intelligence (AI), autonomous vehicles, the Internet of Things, and robotics, increasingly require complex and fast computing. This is because an enormous amount of information must be calculated instantly and results transmitted in real-time to recognize and process objects and events. So far, scientists have focused on improving the performance of central processing units (CPUs) and memory by developing high-speed, large-capacity semiconductors to handle this demand. However, this approach has now reached its limits. Existing von Neumann architecture computers use digital bits, which require sequential processing of one instruction at a time. Since memory and CPU are separate and must exchange information, bottlenecks in data pathways are inevitable. Even the fastest and largest-capacity semiconductors currently available take considerable time to process complex information akin to human sensory perception. Cooling the CPU and memory consumes a lot of electricity, ultimately requiring significant hardware resources. One promising alternative being researched to overcome this is 'neuromorphic' semiconductor technology. The term is a combination of human brain cells (neurons) and the English word 'morphic' (to imitate), meaning it aims to create an 'artificial brain' modeled on the operating principles of the human brain to replace existing semiconductors.

In the fourth game of the 'Google DeepMind Challenge Match,' 9-dan Lee Sedol (right) makes the first move. / Photo by Korea Baduk Association

◇The Paradox of ‘AlphaGo’ vs Lee Sedol

In March 2016, Google DeepMind's AI 'AlphaGo' announced the dawn of the AI era by defeating Lee Sedol 9-dan, then the world's top Go player. Paradoxically, this match also revealed the limitations of existing computing hardware. AlphaGo calculated countless possible moves on the Go board using a 'deep learning' method that mimics human thinking. However, the hardware still used the von Neumann computing architecture established in 1949, i.e., a memory-CPU system based on digital bits and the stored-program concept.

AlphaGo required the connection of about 3,000 enterprise servers to operate. It involved 1,202 CPUs, 176 graphics processing units, 1.03 million memory semiconductors, and over 100 scientists, consuming a massive 170 kW of power per hour. In contrast, although he lost, human Lee Sedol 9-dan only needed his brain and a cup of coffee. Moreover, he fueled his energy with just two bananas. The human brain operates on ultra-low power of about 20 W. While this match heralded the arrival of the AI era, it also clearly exposed the limitations of existing computing architectures. Von Neumann computers sequentially process information using digital bits and have separate storage and processing units, causing structural delays known as the von Neumann bottleneck during data transfer.

◇Imitate the Human Brain

Scientists are researching new computing architectures to overcome these limitations. One approach is 'quantum computing,' which uses quantum physics' 'qubits' to process enormous information instantly. However, commercialization is slow due to physical constraints requiring exclusion of all physical interference such as zero gravity and absolute temperature (?273°C) to realize qubits.

Therefore, neuromorphic technology, which aims to create an 'artificial brain' by mimicking the human brain, has emerged as an alternative. The human brain consists of approximately 100 billion neurons and 100 trillion synapses (connections between neurons). As the saying goes, 'a glance tells all,' the human brain does not calculate each piece of information sequentially like digital bits but instantly perceives all information, performs reasoning, stores, and executes it. The 100 billion neurons and 100 trillion synapses operate in a parallel structure, simultaneously performing memory, calculation, and transmission tasks.

◇A 'Computer' That Understands at a Glance

The brain receives information from the eyes, nose, ears, mouth, and skin, sending signals in all directions to quickly abstract patterns from vast and disordered data, enabling recognition and thought. Neuromorphic semiconductors contain components that imitate neurons and synapses instead of traditional semiconductor 'transistors' or 'cells.' Like the brain, they receive information as event units and simultaneously store, compute, and transmit it. They can perform computation, storage, and learning of various data such as images, videos, and sounds within a single semiconductor. They operate on ultra-low power without generating heat, drastically reducing power consumption. Research using organic materials is also underway, promising conservation of mineral resources like rare earth elements and reducing environmental destruction and pollution.

Professor Choi Yang-gyu of KAIST's Department of Electrical Engineering explained, "Von Neumann computers consist of a CPU for data computation and memory for data storage, causing significant power consumption and time delays due to massive data transfers between CPU and memory. Neuromorphic semiconductors imitate the human neural network structure, enabling parallel and simultaneous computation and storage without separating CPU and memory, improving power efficiency and processing speed."

Spin transistor device

◇Difficult to Implement with Transistors

There are two main approaches to implementing neuromorphic semiconductors. One is to create neurons and synapses as circuits like conventional semiconductors. This can be seen as accelerators or application-specific integrated circuits (ASICs) specially designed for fast AI algorithm processing. Since major semiconductor companies like Intel and IBM are involved, commercialization is expected in about five years.

Intel's neuromorphic chip, Loihi, is considered the most advanced. Developed using a 14-nanometer (nm) process, the Loihi chip consists of 1.3 million digital neurons across 128 cores and 130 million synapses, operating asynchronously. It is known to be up to 1,000 times faster and about 300 times more power-efficient than conventional semiconductors. Universities such as Manchester, Heidelberg, Stanford, and Zurich, along with IBM, are also developing neuromorphic hardware. However, there are limitations: more than 10 transistors are needed to configure a neuron, and at least 6 transistors (SRAM-based) are required for a synapse.

◇Development of Devices Resembling Human Neurons

Therefore, research is underway to develop new devices, not circuit-based, to create second-generation neuromorphic semiconductors that integrate neurons and synapses into a single device. This can reduce area and power consumption, achieving the original goals of ultra-low power and ultra-high-speed computing. In August 2021, Professor Choi and his KAIST research team developed a 'neuromorphic semiconductor module mimicking the human brain' by simultaneously integrating neurons and synapses on the same plane. In February, they developed a neuromorphic module mimicking human tactile neurons, capable of recognizing pressure and outputting spike signals, enabling neuromorphic tactile recognition systems. The Korea Institute of Science and Technology (KIST) also developed core technology for next-generation low-power neuromorphic computing devices using 'skyrmions,' vortex-shaped nano spin structures, in 2020.

Professor Choi said, "Device-based neuromorphic semiconductor technology faces commercialization challenges due to immature processes from new materials, device non-uniformity, and degradation from repeated operations. However, these issues may be resolved by utilizing CMOS device structures and processes, potentially accelerating commercialization." To implement second-generation neuromorphic semiconductors, research is focusing on memristor devices that simultaneously function as memory and variable resistors. Memristors, a combination of memory and resistor, are classified into flash memory, RRAM, PRAM, MRAM, etc., depending on materials and implementation methods.

An academic source stated, "Currently, memristor-based research is the most active, but extensive research is also ongoing on various memory devices. Globally, including Korea, the US, Europe, and China, research is vigorous. Most studies have been at the unit function block level but are recently evolving toward testing implementation possibilities at the system level."