Research On ReRAM-Based Efficient Technology And Simulation For Processing-in-Memory System

Resistive Random Access Memory(ReRAM)is expected to achieve storage class memory due to its excellent performance,good scalability,high density,and low idle power.With its unique crossbar array structure,ReRAM can perform vector-matrix multiplication(VMM)in a time complexity of O(1)by Ohm’s law and Kirchhoff’s current low.The ReRAM and its crossbar structure have made it possible to realize both storage and in-situ computing in the same array,which accomplishes Processing-in-Memory(PIM).Many ReRAM-based PIM architectures for neural networks(NN),graph computing,and various applications are proposed to reduce the data movement overhead in conventional von Neumann architecture.However,limited by non-ideal factors,the current ReRAM-based PIM systems suffer from low efficiency and reliability of storage and computing.In addition,the high overhead of auxiliary logic and the lack of efficient simulation technology also restricts the further promotion of PIM systems.In this thesis,the following three aspects of research are carried out from storage,computing,and simulation in view of the above problems.To overcome the problems of poor performance and high energy consumption in storage caused by non-ideal factors,a reconfigurable hybrid memory technology,called Aliens,based on Complementary Resistive Switch(CRS)is proposed.CRS cell uses two kinds of high resistance state(HRS)to represent 0 and 1.As the cells of the whole array are kept in HRS,the non-ideal sneak current and IR-drop in the array are well restrained to provide better performance and energy efficiency.However,a destructive read is needed to distinguish the two different high resistance states for CRS cells,which brings redundant recovery write delay and energy.A CRS cell can also behave as a typical ReRAM under a different voltage level,which is called the memory(MEM)mode of CRS.Exploiting the dual CRS/MEM mode of CRS cells,a hybrid cell mode organization and a lazy-switch algorithm are proposed.The basic idea of our method is to keep directly accessed cells in the MEM mode while the rest in the CRS mode.The designed cell mode organization not only introduces tiny recording overhead,but also ensures the effective suppression of non-ideal factors.The proposed lazy-switch algorithm delays the recovery write operations for CRS cells and makes use of the locality of memory access to make more read operations hit MEM mode cells.The test results show that compared with the MEM mode only array,Aliens saves 94.8%energy consumption on average and improves the performance by 1.65×.Compared with the CRS mode only array,Aliens achieve a lifetime of 10.7×.To overcome the problems of poor computing accuracy caused by non-ideal factors and huge overhead introduced by auxiliary logic,an efficient ReRAM-based PIM technology for Naive Bayesian algorithm is proposed,called NB Engine,with a carefully designed analog peripheral circuit.The two core operations of the Naive Bayesian algorithm are the probability calculation and the election with the highest probability class.For the probability calculation,a non-ideal-factors considered mapping strategy of the Naive Bayesian algorithm is proposed.By including attribute weights in the conductance of ReRAM cells,the influence of the nonlinear resistance non-ideal factor is eliminated.At the same time,the core operator in the Naive Bayesian algorithm is converted into a dot product pattern which can be mapped to the crossbar array to utilize the high parallel VMM computing ability for implementation.For electing the class with the highest probability,an analog parallel comparison peripheral circuit based on the one-hot code is designed.The designed peripheral only causes slight modification to the conventional read circuit that avoids the enormous auxiliary logic overhead.The test results show that compared with the Naive Bayesian algorithm implemented by the software,Nb Engine introduces an average accuracy decrease by only 1.4%with a maximum performance improvement of 2289.6×.Compared with the conventional NN PIM-compatible design for the Naive Bayesian algorithm,NB Engine saves 96.2%energy consumption and 45.2%chip area.To overcome the problems of poor performance and huge overhead of the comprehensive array model and inaccurate simulation results of the approximate array model,an iterative refinement model,called IR2S,is proposed with the corresponding simulation algorithm which guarantees exact simulation results.The existing works either adopt a complex and time-consuming comprehensive model or a simplified but inaccurate approximate model.Our proposed IR2S model ensures exact simulation results under the convergence condition.For an array of size n×n,IR2S algorithm realizes a spatial complexity of O(n~2)and a time complexity of O(n~2log_γδ)to achieve<0.01%maximum relative error.Furthermore,an IR-drop considered initialization method is proposed to set a more accurate initial node voltage to reduce the number of iterations.A modeling framework Pim Torch is designed to simulate applications running in the PIM system.The test results show that IR2S algorithm only requires 6 iterations to achieve less than 0.01%maximum relative error,while the approximate model has a maximum relative error of 2.6%.Compared with HSPICE comprehensive model,IR2S improves performance by 3389-36667×.In addition,IR2S can be further accelerated by multi-thread or GPU.