Write-Energy Relaxation of MTJ-Based Quantized Neural-Network Hardware

This paper evaluates WKH bit-error tolerance of quantized neural networks (QNNs) for energy-efficient artificial intelligence (AI) applications utilizing stochastic properties of magnetic tunnel junction (MTJ) devices. Since QNNs have potentially high bit-error tolerance, they do not require large write currents to guarantee the certainty of the information held in the MTJ devices. By artificially adding bit errors to their weights, it is demonstrated that QNNs with binarized data representation achieve better error tolerance than any other ones in terms of the degradation rate of the recognition accuracy. In addition, based on the evaluation results, we show the possibility of reducing the write energy of MTJ devices up to 42% by exploiting high bit-error tolerance of the binarized QNN.