Research And Application Of IN-MRAM Bitwise Computing

In the traditional von Neumann architecture,data transfer between processor and memory consumes high energy and deteriorates the system performance.This phenomenon is called“memory wall”.Computing in memory(CIM)breaks through this bottleneck and realizes the fusion of memory and processor.CIM based on magnetic random access memory(MRAM)shows extremely high advantages in terms of memory access speed and power efficiency,which is a promising candidate for energy-efficient computing operations in artificial intelligence edge devices.MRAM-based CIM mainly relies on the reconfiguration of the sense or write circuit.However,circuit design for MRAM-based CIM imposes a number of challenges,including device physical parameter variations,limited read margin and the area-latency-energy tradeoff for multibit operations.In order to further study the feasibility of MRAM-based CIM design,this thesis explores the sensing and writing based CIM respectively.Based on Spin-Transfer Torque(STT)MRAM,firstly,a cycle-sensing margin enhancement(CSME)scheme is proposed to increase the input voltage diffference and expand the sensing window through periodic charging and discharging,thereby improving the reliability of the sensing-based CIM.Secondly,a self write termination method is designed to avoid the redundant energy consumption,and five Boolean logic(invert,AND,OR,XOR,full adder)can be realized with little configuration.Then a read-verify-write circuit combined with reverse bias based on magnetoelectric random access memory is designed to improve the magnetic stability of the writing operation,and an XOR logic operation based on voltage-controlled switching is demonstrated.Additionally,two convolution calculation circuits are realized,including binary-input ternary-weighted(BITW)convolution calculation network and binary neural network(BNN).BITW network is realized by a combination of sensing and writing,while BNN is realized by sensing only.Using TSMC 28 nm CMOS process,read with CSME scheme realizes a minimum 2.4×margin improvement and 14.1% read error rate reduction at 70% tunnel magnetoresistance ratio(TMR),under 0.6V supply voltage comparing to traditional voltage sense amplifier(VSA).The self-write-termination method achieves 84.7% writing energy saving within 20 ns write access duration.A high-level simulation is performed with 28×28 pixels image similarity analysis.It demonstrates 24% dynamic energy reduction comparing to the previous method.The read-verify-write circuit has achieved a 52%-68.7% write error rate reduction compared with the traditional design,and the voltage-controlled writing-based CIM achieves less than 3ns delay,which is similar to the sensing-based CIM.Finally,the image edge processing is realized by using BITW network under TSMC 28 nm process.Under HK 28 nm,The BNN circuit can achieve within 20 ns,and the power is less than 2.74 p J.Sensing-based CIM can achieve fast speed and low power,but limited by TMR,the number of operations for this computing is limited.Writing-based CIM is more suitable for the scenario for higher accuracy and smaller area.