Radiation-Hardened Processing-In-Memory Crossbar Array With Hybrid Synapse Devices for Space Application

This paper presents a multilayer perceptron (MLP) that offers excellent accuracy for classifying MNIST handwritten images considering radiation-induced bit failures. By introducing a stochastic model for radiation effect on ideal error-free MLP, the performance degradation of the neural network on space application is inevitable. Radiation-hardened processing in memory (PIM) should be developed with minimum hardware additives to utilize edge devices more practically in space. In the previous studies on digital synaptic devices to overcome radiation-related side effects, as the number of transistors in the unit storage device increases, more tolerance to radiation is expected. However, when all weight devices are replaced with bulky ones, the overall volume of the processor increases. This work proposes a digital hybrid synaptic device that only uses a larger device on the most significant bit (MSB) when the radiation effect is considered. With minimum hardware overhead for synapses, improved performance in the classification of MNIST is obtained. From the Neurosim framework with a single hidden layer, the accuracy is dramatically improved while sacrificing 1-bit weight information.