Pj-AxMTJ: Process-in-memory with Joint Magnetization Switching for Approximate Computing in Magnetic Tunnel Junction

Abstract

In order to realize high efficient magnetization switching in magnetic tunnel junction (MTJ), several potential alternative mechanisms have been realized to replace spin transfer torque (STT) method, such as the STT-assisted precessional voltage controlled magnetic anisotropy (VCMA), and the spin orbit torque (SOT) erasing plus STT programing. In this paper, we propose a method denoted as the Process-in-memory with Joint magnetization switching for Approximate computing in Magnetic Tunnel Junction (Pj-AxMTJ), by using the 1T-1M and 3T-1M bit-cell structures. The proposed method aims to implement a low-precision computational memory with dynamic approximate computing. Specifically, four nonvolatile approximate full adders (AxFAs) are proposed based on the writing operations of different types of magnetic random access memory. As no peripheral circuits but the memory bit-cells are used in the proposed design, the resultant area is significantly small. Moreover, the AxFAs can be easily reconfigured into memory units with simple wire connections. The simulation results for the proposed designs are then presented to show the precision-power-area-speed tradeoffs for addition operation. Finally, the accuracy of the AxFAs are further evaluated in an image processing application.