A Low-Power High-Speed Spintronics-Based Neuromorphic Computing System Using Real-Time Tracking Method

Abstract

In spintronic-based neuromorphic computing systems (NCSs), the switching of magnetic moment in a magnetic tunnel junction (MTJ) is used to mimic neuron firing. However, the stochastic switching behavior of the MTJ and process variations effect lead to a significant increase in the stimulation time of such NCSs. Moreover, current NCSs need an extra phase to read the MTJ state after stimulation, which is in contrast with real neuron functionality in human body. In this paper, the read circuit is replaced with a proposed real-time sensing (RTS) circuit. The RTS circuit tracks the MTJ state during stimulation phase. As soon as switching happens, the RTS circuit terminates the MTJ current and stimulates the post neuron. Hence, the RTS circuit not only improves the energy consumption and speed, but also makes the operation of the NCS similar to real neuron functionality. The simulation results in 65-nm CMOS technology confirm that the energy consumption and speed of the proposed RTS-based NCS are improved at least by 40% and 2.22X compared with a typical NCS, respectively. Finally, utilizing the RTS-based NCS in image processing applications, such as character recognition and edge detection, can lead to 90.3% improvement in energy delay products compared with the typical NCS.