Design Of Storage-compute Fusion Circuit For Spin Magnetic Memory For Convolutional Neural Network

In data-intensive applications,the memory wall caused by traditional Von-Neumann architecture will reduces the efficiency of the system.Computing-in-Memory(CIM)is taken as a promising approach to solving the above problem by completing calculations in memory.However,the area overhead of traditional MRAM based CIM is large and the result can't be stored in situ.Approximate computing(AC)is another method to reduce power consumption,area,and delay of circuits,which is suitable for accuracy-insensitive applications such as image and video processing.MARM can further these AC advancements in power savings and area reduction due to its near-zero current leakage.A MRAM based in-memory-Boolean logic was realized with a novel writing mode by using Toggle-Spin-Torques-MRAM bit-cell to solve the problem of traditional MRAM based CIM.Then,a variety of computing circuits such as adders,4-2 compressors and multipliers based on voter logic were proposed,and the corresponding approximate circuits were designed to reduce the area and save power.Finally,the TST-MRAM multiplication array for neural network was proposed,and each module of the circuit was introduced separately.The proposed approximate adder has the smallest area compared to the traditional approximate adders based on CMOS and other non-volatile memory.The proposed approximate 4-2 compressors show reduced power consumption and delay than previous compressors implementations by more than 48.4% and 50%.With SMIC 28-nm process,the area consumption of TST-MRAM based approximate multipliers were reduced by more than 70% compared with the traditional CMOS multipliers.The structural similarity(SSIM)of images sharpened by using proposed multipliers were greater than 95%,and the peak signal-to-noise ratio(PSNR)of these images were beyond 30 d B.Compared with multipliers implemented by conventional logic,the proposed multipliers have fewer steps and smaller area,making it suitable for accuracy-insensitive applications.