Samsung Electronics today announced that it has successfully demonstrated the world's first MRAM (Magnetoresistive Random Access Memory)-based memory computing. A paper on this innovation was published on the Nature website on January 12 and will be published in print in the forthcoming journal Nature.
From left to right: Dr. Donhee Ham, Dr. Seungchul Jung and Dr. Sang Joon Kim
The paper, titled "A crossbar array of magnetoresistive memory devices for in-memory computing," demonstrates Samsung's leadership in memory technology and its efforts to converge memory and system semiconductors for next-generation artificial intelligence (AI) chips.
The research was led by the Samsung Institute for Advanced Technology (SAIT) in close collaboration with Samsung's electronics foundry business and semiconductor R&D center. The paper's first author, Dr. Seungchul Jung, a SAIT employee researcher, and co-corresponding authors Dr. Donhee Ham, a SAIT researcher and professor at Harvard University, and Dr. Sang Joon Kim, vice president of SAIT technology, spearheaded the study.
In a standard computer architecture, data is stored in a memory chip and data calculations are performed in a separate processor chip. In contrast, in-memory computing is a new computing paradigm that attempts to do both data storage and data computation in an in-memory network. Because this scheme can handle large amounts of data stored in the memory network itself without having to move the data, and because data processing in the memory network is performed in a highly parallel manner, power consumption is greatly reduced. As a result, in-memory computing has become one of the promising technologies for enabling the next generation of low-power AI semiconductor chips.
There has been a lot of research on in-memory computing around the world. In terms of memory computing, the current development directions are mainly RRAM (resistive random access memory) and PRAM (phase change random access memory). In contrast, despite the speed of operation, durability, and mass production of MRAM, it has so far been difficult to use MRAM (another type of non-volatile memory) for in-memory computing. This difficulty comes from mram's low resistance, which does not enjoy the benefits of reduced power consumption when MRAM is used in standard memory computing architectures.
Researchers at Samsung Electronics have provided solutions to this problem through architectural innovations. Specifically, they successfully developed an MRAM array chip that demonstrates memory computing, solving the problem of small resistance of a single MRAM device by replacing the "current-sum" of the traditional standard memory computing architecture with a new "resistance sum" resistor.
Subsequently, Samsung's research team tested its performance by running this MRAM memory computing chip to perform ARTIFICIAL computing. The chip achieves 98 percent accuracy when classifying handwritten numbers and 93 percent when detecting faces from scenes.
The researchers also propose that this new MRAM chip can not only be used for in-memory computing, but also as a platform to download a network of biological neurons. This is in line with the neuromorphic electronic vision set out by Samsung researchers in a recent opinion paper published in the September 2021 issue of the journal Nature Electronics.
Dr Seungchul Jung, lead author of the paper, said: "Memory computing has similarities with the brain, because in the brain, computation also takes place in the network of biological memory, that is, synapses, which are the points where neurons come into contact with each other. In fact, while the calculations currently performed by our MRAM networks have a different purpose than those performed by the brain, this solid-state memory network may be used as a platform in the future to mimic the brain by mimicking the brain's synaptic connections."