▲(From left) Samsung Electronics Fellow Ham Don-hee, Research Fellow Jeong Seung-cheol, and Master Kim Sang-jun



Next-generation low-power AI chip attracts attention as a promising technology
Expected to help with future neuromorphic technology, research, and development

Samsung Electronics has expanded the horizon of next-generation low-power AI chip technology by successfully achieving a low-power design that overcomes the limitations of existing magnetoresistive memory (MRAM).

Samsung Electronics announced on the 13th that its research team has implemented in-memory computing based on MRAM for the first time in the world.

Existing computers are configured with separate memory chips, which are responsible for storing data, and processor chips, which are responsible for performing data operations.

In-memory computing is a cutting-edge chip technology that not only stores data within memory, but also performs operations on the data.

Since it performs parallel calculations within memory without moving a large amount of information in memory, power consumption is significantly reduced, and it is attracting attention as a promising technology for creating next-generation low-power artificial intelligence (AI) chips.

The implementation of in-memory computing using non-volatile memory such as RRAM (resistive random access memory) and PRAM (phase-change memory) has been a topic of research interest worldwide for the past several years.

Although MRAM has the advantages of high data stability and high speed, its low resistance value does not provide much power advantage when applied to in-memory computing, so it is not implemented as in-memory computing. Couldn't do it.

Samsung Electronics' research team succeeded in designing a low-power memory by proposing an in-memory computing structure using a new concept of 'resistance summation' rather than the existing 'current summation' method to overcome these limitations of MRAM.

The researchers applied the performance of MRAM-based in-memory computing chips to artificial intelligence calculations.

It has been verified to operate with an accuracy of up to 98% in digit classification and 93% in face detection.

This study is of great significance in that it is the world's first in-memory computing implementation of MRAM that can be mass-produced by combining it with system semiconductor processes, and expands the horizon of next-generation low-power AI chip technology.

The research results were published in the world-renowned academic journal 'Nature' on the 12th (UK local time).

This study was led by Senior Researcher Jeong Seung-cheol of the Samsung Electronics Advanced Institute of Technology, with fellow Ham Don-hee of the Advanced Institute of Technology, professor at Harvard University, and master Kim Sang-jun of the Advanced Institute of Technology participating as co-corresponding authors.

Researchers from Samsung Electronics' Advanced Institute of Technology, Semiconductor Research Institute, and Foundry Business Division also jointly participated in the research.

Jeong Seung-cheol, a senior researcher at Samsung Electronics’ Advanced Institute of Technology, said, “In-memory computing is a technology that combines memory and computation, and has some similarities to the human brain, where memory and computation are intertwined.” He added, “This research could also be helpful in the future research and development of neuromorphic technology that mimics the actual brain.”