Research On Accuracy Recovery Of RRAM-Based Neuromorphic Computing Systems

The RRAM-based crossbar structure is one of the most potential architectures to implement neuromorphic computing systems,due to its ability to support energy-efficient Processing-In-Memory(PIM).However,as the intrinsic device characteristics and immature manufacturing process of RRAM,there exist some non-ideal factors in RRAM cells and RRAM Crossbars.These non-ideal factors will unexpectedly affect the operands during computation,thus cause system accuracy degradation.To achieve higher system accuracy,this thesis selects the digital RRAM device(two states per cell)with higher cell stability,and proposes specified computation accuracy recovery schemes of neuromorphic computing systems based on digitalRRAM Crossbars.During computation processes,the digital-RRAM Crossbar mainly suffers from resistance drift caused by high temperature as well as IR drop caused by interconnect resistance.On the one hand,resistance drift will affect the weight data of the neural network stored in Crossbars.In order to mitigate the impact of the high-temperature environment of Crossbars in the 3D stacked neuromorphic computing architecture on the accuracy,Weight Significance based Remapping scheme(WS-R)is proposed.WS-R classifies the weight matrix according to the weight significance that is determined by the digit position and the proportion of effective data.Moreover,WS-R divides the Crossbar into cold and hot areas based on the temperature distribution.Then,a remapping step is used to avoid the high significance weights,which contribute a lot to network outputs,from being mapped to the hot areas.Thus,the impact of high temperature on the computation accuracy is reduced.The experimental results show that,compared with the remapping scheme based on SWV metric,WS-R improves the computation accuracy by 37.8%-56.5%,under the weight precision of 4-8 bits.On the other hand,IR drop caused by interconnect resistance in the Crossbar will reduce the value of the input voltage vector.In order to compensate for the interference of IR drop,Location and Data Pattern based IR drop Compensation scheme(LDP-IRC)is proposed,which includes resistance adjustment and drive voltage tuning.The experimental results show that,the deviation of output current of each column is smaller than ±10%,when using LDP-IRC.At the computing system level,LDP-IRC restores the system accuracy close to the ideal value under different network sizes.Finally,the designs to seamlessly integrate WS-R with LDP-IRC are proposed to mitigate the effects of above-mentioned two non-ideal factors at a time.The experimental results show that under the weight precision of 4-8 bits,the system with the combined scheme can achieve up to 73.48%-80.02% computation accuracy,when compared with the ideal situation.