Energy-Efficient CNN Accelerator Using Voltage-Gated DSHE-MRAM

Abstract

Modern convolutional neural networks (CNNs) architectures need millions of parameters to be stored and computed in hardware for image classification. This needs a huge amount of memory and power in conventional hardware accelerator architectures. The concept of neuromorphic and in-memory computing (IMC) using emerging memory technologies has led to energy-efficient computations for hardware accelerators. This work presents a voltage-gated dual-bit spin Hall effect (VG-DSHE) magnetic random access memory (MRAM) device-based accelerator. The VG-DSHE-MRAM provides efficiency in terms of power and speed as compared to other MRAM devices. A crossbar array is implemented using VG-DSHE devices to exploit high-density storage, energy-efficiency, and fast multiply and accumulate (MAC) computation. Finally, a complete hardware implementation of CNN architecture is presented for image classification using the CIFAR-10 dataset without any significant accuracy degradation. The proposed VG-DSHE-based CNN accelerator is $2.1\times $ more energy-efficient as compared to conventional differential spin Hall effect (DSHE)-based designs and achieves a throughput efficiency of 1.57, 0.49, 0.035, and 0.018 TSOPS/W for VGG8, VGG16, AlexNet, and ResNet18 architectures, respectively. Furthermore, the proposed accelerator is compared with other emerging memories for AlexNet architecture that shows $181\times $ , $14.76\times $ , $2.4\times $ , $2.25\times $ , and $2.1\times $ improvement in crossbar power consumption as compared to phase change memory (PCM), resistive random access memory (RRAM), spin transfer torque (STT), spin-orbit torque (SOT), and DSHE, respectively.