Accuracy rate 98%! Samsung's world premiere MRAM magnetoresistive in-memory computing

Samsung Semiconductor announced that through structural innovation, it has realized MRAM (magnetoresistive random access memory) based on in-memory computing (In-Memory Computing), further expanding the frontier of Samsung's next-generation low-power artificial intelligence chip technology.

In the traditional computing system, the data in memory is transferred to the data calculation unit of the processing chip for processing, and the requirements for bandwidth and latency are very high.

In-memory computing is a new computing mode, which can also be called "memory-computing integration", which simultaneously performs data storage and data computing processing in memory, without the need to move data back and forth.

At the same time, data processing in the memory network is performed in a highly parallel manner, so improving performance while greatly reducing power consumption.

Compared with other memories, MRAM magnetoresistive memory has obvious advantages in terms of running speed, life, and mass production, and the power consumption is much lower than that of traditional DRAM, the key is that it also has non-volatile characteristics, that is, power failure will not lose data.

However, MRAM magnetoresistive memory has been difficult to use for in-memory computing because it cannot take advantage of low power consumption in standard in-memory computing architectures.

The Samsung research team designed a new in-memory computing architecture called "resistance sum" to replace the standard "current-sum" architecture and successfully developed a MRAM array chip that can demonstrate the in-memory computing architecture, named "crossbar array of magnetoresistive memory devices" for in-memory computing for in-memory computing)。

This array successfully solves the problem of small resistance of a single MRAM device, thereby reducing power consumption and enabling MRAM-based in-memory computation.

According to Samsung, when performing AI calculations, MRAM in-memory calculations can achieve a 98% success rate of handwriting recognition and a 93% accuracy rate of face recognition.