Energy Efficient DSHE based Analogue Multiply Accumulate Computing Crossbar Architecture

The implementation of non-volatile memories (NVMs) based in-memory computing (IMC) have a great potential for neural network applications. Both digital and analog based IMC approaches have been explored for multiply accumulate (MAC) operations that are extensively used in deep convolutional neural networks. Various NVMs, including resistive memory, phase-change memory and magnetic random access memory (MRAM) have been used for such implementation using crossbar array architectures. However, a major challenge with the MRAM crossbar array is due to the low resistance of synaptic devices resulting in large power consumption in a conventional crossbar array that uses current summation for analogue MAC operations. Differential spin Hall effect (DSHE) MRAM is the most suitable candidate among other MRAM devices due to its high storage density, high speed, and energy efficient operation. This work presents a DSHE MRAM-based crossbar array to overcome the high power consumption issue by using resistance summation for analogue MAC operations. The proposed approach achieves 11.1x and 1.83x more energy-efficient MAC computation compared to the conventional spin transfer torque (STT) and spin orbit torque (SOT)- current summation based methods, respectively. Moreover, the presented DSHE-based approach has 11.05x lower energy consumption than the STT-based resistance summation approach for MAC computation. The proposed approach has 2.42x and 2.5x improvement in latency compared to the current summation and resistance based STT-MRAM, respectively.