MRAM-based BER resilient Quantized edge-AI Networks for Harsh Industrial Conditions

Abstract

We investigate Edge-AI Inference (EAI) architectures based on 22nm FD-SOI embedded-MRAM (eMRAM) using quantized neural networks (QNN) for inference applications in harsh industrial conditions having strong magnetic field and wide operating temperature (-40∼125 °C). We achieved best case test accuracy of 98.99% with Quantized-Convolutional Neural Network (QCNN) and 89.94% with Quantized-Multi-layer Perceptron (QMLP) surpassing prior reported results in literature on MNIST dataset. By exploiting BER resilience of QNN, we show that eMRAM based EAI offers a superior magnetic immunity of ≈ 700 Oe at 125 °C (≈ 98% accuracy) without the use of ECC and significant energy saving of ≈ 14% for QCNN and ≈ 11% for QMLP. A detailed analysis on the tradeoff between retention time, write energy and inference accuracy is presented.