Error-Sensitivity-Aware Write-Energy Optimization for an MTJ-Based Binarized Neural Network.

This paper proposes a write-energy optimization method in the processing of binarized neural networks (BNNs) utilizing the stochastic properties of magnetic tunnel junction (MTJ) devices for energy-efficient artificial intelligence (AI) applications. By exploiting the high bit-error tolerance of BNNs, the performance requirements for writing to MTJ-based on-chip buffer memory can be relaxed, resulting in the realization of an energy-efficient BNN. Furthermore, by utilizing the fact that each layer of the BNN has a different sensitivity to bit errors, it is possible to optimize the write energy on a layer-by-layer basis. Through the evaluation results for BNNs with various error conditions, it is demonstrated that the proposed method reduces write energy for MTJ-based on-chip buffer memory by 14% compared with the conventional method.