Samsung shows MRAM can support in-memory computing
The use of crossbar arrays of memory to perform analog multiply-accumulate (MAC) operations has been demonstrated for resistive memory, phase-change memory and flash memory. MRAM is being adopted as an embedded non-volatile memory in leading-edge manufacturing processes for its endurance but its low resistance would result in large power consumption in a conventional crossbar array that uses current summation for analog MAC operations.
In-memory computing has emerged as one of the promising technologies to realize next-generation AI semiconductor chips because it allows local processing of data stored in memories and can minimize the movement of data thereby saving power consumption.
Engineers from Samsung have authored a paper, ntitled ‘A crossbar array of magnetoresistive memory devices for in-memory computing’ in Nature that reports on a 64 by 64 crossbar array based on MRAM cells that overcomes the low-resistance issue with an architecture that uses resistance summation for analog MAC operations.
The array is integrated with readout electronics in 28nm CMOS. Using this array, a two-layer perceptron is implemented to classify 10,000 Modified National Institute of Standards and Technology (MNIST) digits with an accuracy of 93.23 per cent (software baseline: 95.24 per cent). In an emulation of a deeper, eight-layer visual geometry group-8 neural network with measured errors, the classification accuracy improves to 98.86 per cent (software baseline: 99.28 per cent). We also use the array to implement a single layer in a ten-layer neural network to realize face detection with an accuracy of 93.4 per cent.
Next: Combined research
The research was led by Samsung Advanced Institute of Technology (SAIT) in close collaboration with Samsung Electronics Foundry Business and Semiconductor R&D Center.
The researchers have suggested that not only can this MRAM chip be used for in-memory computing, but it also can serve as a platform to download biological neuronal networks. This is along the line of the neuromorphic electronics vision that Samsung’s researchers recently put forward in a perspective paper published in the September 2021 issue of the journal Nature Electronics (see Researchers want to ‘copy and paste’ the brain into memory).
“In-memory computing draws similarity to the brain in the sense that in the brain, computing also occurs within the network of biological memories, or synapses, the points where neurons touch one another,” said Seungchul Jung, the first-nmaed author of the paper. “In fact, while the computing performed by our MRAM network for now has a different purpose from the computing performed by the brain, such solid-state memory network may in the future be used as a platform to mimic the brain by modelling the brain’s synapse connectivity.”
Related links and articles:
Samsung presents AI processing-in-memory options
Researchers want to ‘copy and paste’ the brain into memory
Samsung embeds AI accelerator in memory chip
Unsupervised in-memory AI learning goes digital
Processor-in-memory DRAM benchmarked on Xeon server