Magnetoresistive RAM can replace conventional dynamic memory with many advantages. However, when it comes to computing directly in memory, a number of problems arise.
As pointed out in a scientific paper by scientists from Samsung, the development of a magnetoresistive random access memory array with spin momentum transfer remains a very laborious task, despite the practical advantages of this new technology. The difficulty is due to the low resistance of the MRAM, which would result in high power consumption in a conventional die array using current summation for analog multiply-accumulate operations. The company has been able to create a 64x64 jumper array based on MRAM cells that can overcome the low resistance problem with an architecture that uses resistance summation for analog multiply-accumulate operations. This array is integrated with readout electronics based on 28nm CMOS technology. With this array, a two-level perceptron was implemented to classify 10,000 handwritten digits with 93.23% accuracy. When emulating a deeper eight-layer Visual Geometry Group-8 neural network with measured errors, the classification accuracy increases to 98.86%. Samsung also used an array to implement one layer in a ten-layer neural network, and a high accuracy of 93.4% was achieved in the face recognition task.
Samsung's development could allow these MRAM computations to be used to create AI, including as a brain simulation platform by mimicking synaptic connections.