Samsung announced the successful demonstration of the world’s first in-memory computing based on MRAM, a magnetoresistive random-access memory. An article about this was published on January 12 in the online version of Nature. Samsung said the achievement demonstrates its leadership in memory technology and its efforts to integrate memory and computing chips for next-generation artificial intelligence chips.
In standard computer architectures, data is stored in memory chips and calculations are performed by the central processing unit. In-memory computing is a new computing paradigm in which the memory subsystem not only stores data, but also works with them. Since this approach allows processing large amounts of data without having to move it from the memory subsystem, and this processing is performed in a highly parallel manner, power consumption is significantly reduced compared to traditional systems. Thus, in-memory computing is one of the promising technologies for realizing the next generation of chips for AI operation, which can boast of minimal power consumption.
For this reason, research in the field of memory computing is being actively conducted all over the world. RRAM (Resistive Random Access Memory) and PRAM (Phase Change Random Access Memory) and MRAM were used to demonstrate them. The latter has so far been rather difficult to use for in-memory computing, despite all its advantages, such as speed, durability, and the fact that it is produced on a large scale. These difficulties were due to the low resistance of MRAM, due to which it cannot provide a reduction in power consumption when used in a standard memory computing architecture.
Samsung researchers have come up with architectural innovations that could solve the problem. They succeeded in developing an MRAM chip for memory computing by replacing the standard “current sum” computing architecture with a new “resistance sum” architecture that solves the low resistance problem of individual MRAM devices.
The Samsung research team has tested the new solution in action. MRAM calculations have been tested with artificial intelligence operations. The chip achieved an accuracy of 98% in handwritten digit recognition and 93% in face detection in scenes.
The researchers note that the use of MRAM, the production of which has already reached commercial scale, for memory computing expands the possibilities for creating next-generation artificial intelligence chips with low power consumption.
The researchers suggest that the MRAM chip they have developed can be used not only for memory computing, but also as a platform for loading biological neural networks. According to them, computation in memory is similar to brain activity, since in the brain, computation occurs at a network of synapses – points where neurons touch each other. Thus, a fresh development, in theory, can be used as a platform for simulating the brain by modeling synapses.
If you notice an error, select it with the mouse and press CTRL + ENTER.