Samsung today announced the successful demonstration of the world’s first in-memory computing based on MRAM, a multi-resistive random access memory (MRAM). An article on this was published in the online edition of Nature on January 12th. Samsung said the success demonstrates its leadership in memory technology and its efforts to combine memory and computer chips for next generation artificial intelligence chips.
In standard computer architectures, data is stored in memory chips and calculations are carried out by a central processing unit. Computing in Memory is a new computing paradigm in which the storage subsystem not only stores data, but also works with it. Since this approach allows you to process large amounts of data without having to move it out of the storage subsystem, and this processing takes place in a highly parallel manner, the power consumption is significantly reduced compared to conventional systems. Thus, in-memory computing is one of the most promising technologies for the implementation of next-generation AI chips that are characterized by minimal power consumption.
For this reason, research in the field of in-memory computing is actively pursued worldwide. To demonstrate it, RRAM (Resistive Random Access Memory) and PRAM (Phase Change Random Access Memory) and MRAM were used. Despite all the advantages such as speed, durability and mass production, the latter has so far been difficult to use for in-memory computing. These difficulties have been with the low impedance MRAM, because of which it cannot offer power savings when used in a standard in-memory data processing architecture.
Samsung researchers have come up with architectural innovations that could solve the problem. They succeeded in developing an MRAM chip for in-memory computing that replaces the standard “running sum” architecture with a new “sum-of-resistance” architecture that addresses the problem of the low impedance of individual MRAMs. Components solves.
The Samsung research team tested the new solution in action. The MRAM computation was tested with artificial intelligence operations. The chip achieved an accuracy of 98% for handwritten digit recognition and 93% for face recognition in scenes.
The researchers find that the use of MRAM, which is already at commercial scale, for in-memory computing expands the possibilities for developing next-generation AI chips with low power consumption.
The researchers suggest that the MRAM chip they developed can be used not only for in-memory computations, but also as a platform for loading biological neural networks. Computing in memory is similar to the activity of the brain, they said, because in the brain, computing takes place at a network of synapses – the points where neurons meet. Thus, the fresh development can theoretically be used as a platform for simulating the brain by simulating synapses.