For in-memory computing, researchers have developed a magnetoresistive random-access memory (MRAM) array chip that can also serve as a platform for loading biological neural networks.
In a standard computer architecture, data is stored in memory chips, and data calculations are performed in separate processor chips. In-memory computing can realize next-generation low-power semiconductor artificial intelligence chips because data storage and computing are possible in the memory network. This leads to a significant reduction in energy consumption.
Samsung’s MRAM array chip demonstrated in-memory computing with a new “sum of resistance” architecture that solved the problem of low resistance. In an AI performance test, the chip was 98% accurate in classifying handwritten digits and 93% accurate in face detection.
Samsung said the research pushes the boundaries of next-generation low-power AI chip technology.
“Computing in MRAM is similar to the brain in the sense that in the brain, computation also takes place in a network of biological memories or synapses, the points where neurons touch each other,” said Jung Seung-chul, a researcher at Samsung.
“While the computations performed by our MRAM network currently have a different purpose than the computations performed by the brain, such a solid-state memory network could in the future be used as a platform to simulate the brain by modeling brain synapse connections,” the scientist added.