Samsung achieves first success with in-memory computing on MRAM

In-memory computing is a rapidly emerging phenomenon that should make data-intensive AI workloads more energy efficient. To do this, certain tasks are performed in memory to save bandwidth and processor time. The technology has already been tested with most available types of non-volatile memory and in dram. At the end of 2021, Alibaba also developed a ram chip with in-memory processing.

Magnetoresistive random access memory or mram fulfills many of the desirable properties for in-memory processing such as speed, durability, non-volatility and the possibility of relatively cheap mass production. However, mram was previously unsuitable due to its low internal resistance. As a result, there is little room for a reduction in energy consumption, which must be one of the major advantages.

Researchers of Samsung now say they have successfully implemented in-memory computing on mram. An array of 64 by 64 cells is used for this. The low resistance problem is overcome by an architecture that uses adding resistors for analog multiplication and addition operations. The mram cells and the necessary readout circuits were produced on a 28 nanometer cmos-transistor process.

That is what the three scientists write in the paper “A crossbar array of magnetoresistive memory devices for in-memory computing” published on January 12 by Nature. The Samsung Advanced Institute of Technology (SAIT) team has tested the new solution in collaboration with Samsung’s Electronics Foundry Business and Semiconductor R&D Center. The AI ​​performance looked promising, with a 98 percent accuracy in recognizing handwritten numbers. The accuracy for facial recognition is 93 percent.

Samsung wants to further explore this technology with the aim of achieving leadership in memory technology and processor development. It is not yet possible to estimate whether and when in-memory computing will come to consumer electronics such as PCs, tablets and smartphones.

In addition to in-memory processing, according to the experts, this mram solution can also be used as a platform for neural networks, which is explained as follows: “In-memory computing has similarities with the brain in that in the brain calculations also take place in the brain. network of biological memory, or synapses, the points where neurons meet.”

Source: Tom’s Hardware


Source: Hardware Info Compleet by nl.hardware.info.

*The article has been translated based on the content of Hardware Info Compleet by nl.hardware.info. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!