Samsung Researchers Develop Advanced Brain-Inspired Chip Technology... Published in Nature (Comprehensive)

In-Memory Computing for Data Storage and Processing Based on MRAM
Expanding Horizons of Low-Power Artificial Intelligence (AI) and Neuromorphic Chip Technologies

Seungchul Jeong, Senior Researcher at Samsung Advanced Institute of Technology, who first implemented MRAM-based in-memory computing in the world and published the research results in Nature (from left), Donhee Ham, Fellow at the Advanced Institute of Technology and Professor at Harvard University, and Sangjun Kim, Master at the Advanced Institute of Technology
[Photo by Samsung Electronics]

[Asia Economy Reporter Kim Heung-soon] Samsung Electronics researchers have become the first in the world to implement semiconductor technology capable of performing memory and computation within a single chip, similar to the human brain, and published their research results on the 13th (Korean time) in the international academic journal 'Nature.' This is expected to accelerate the development of next-generation artificial intelligence (AI) semiconductors that process large volumes of data as well as 'neuromorphic technology' that mimics the human brain.

This research was led by Seungcheol Jeong, a senior researcher at Samsung Electronics' Advanced Institute of Technology, as the first author, with Donhee Ham, a fellow at the Advanced Institute of Technology and professor at Harvard University, and Sangjun Kim, a master researcher at the Advanced Institute of Technology, as co-corresponding authors. Researchers from Samsung Electronics' Advanced Institute of Technology, Semiconductor Research Center, and Foundry Division also participated jointly in the study.

The memory technology implemented by Samsung Electronics researchers is 'in-memory computing' based on magnetic random-access memory (MRAM) devices. This is cutting-edge chip technology that performs both data storage and computation within a single memory unit.

Typically, computers separate memory chips responsible for data storage and processor chips responsible for computation. The central processing unit (CPU) sequentially fetches instructions from memory, executes them, and stores the results back into memory. When the amount of data exchanged between the CPU and memory increases, processing delays occur.

In contrast, applying in-memory computing technology allows parallel computation within the memory itself without data movement, significantly reducing power consumption. Consequently, it is attracting attention as a technology capable of creating next-generation low-power AI chips.

Non-volatile memory devices applicable to in-memory computing include resistive RAM (RRAM), phase-change RAM (PRAM), and MRAM. Previously, in-memory computing technology had been implemented using RRAM and PRAM. Samsung Electronics researchers are the first to achieve this using MRAM.

MRAM is a non-volatile memory with high data stability and fast speed, enabling mass production when integrated with system semiconductor processes. However, due to its low resistance characteristics, applying it to in-memory computing had limitations in power efficiency. Samsung Electronics researchers succeeded in low-power design by proposing a new concept of 'resistance summation' in-memory computing architecture instead of the conventional 'current summation' method.

Samsung Electronics stated, "The significant achievement of this research is the world's first implementation of in-memory computing using MRAM, a non-volatile memory capable of mass production through integration with system semiconductor processes, thereby expanding the horizon of next-generation low-power AI chip technology."

AI semiconductors, which execute large-scale computations such as learning and inference with high performance and low power consumption, are expanding demand across all industries including mobile, automotive, and home appliances beyond servers and clouds. Market research firm Gartner forecasts the AI semiconductor market to grow from $33.148 billion (approximately 39 trillion KRW) in 2020 to $76.77 billion (approximately 91 trillion KRW) by 2025. The industry aims for neuromorphic semiconductor technology that mimics brain neural networks to operate similarly to the human brain in cognition and inference as the ultimate goal.

Senior researcher Jeong said, "In-memory computing is similar to the human brain where memory and computation are intertwined," and added, "This research is expected to contribute to future research and development of neuromorphic technology that truly mimics the brain."

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