The Research On Convolutional Neural Network Accelerator Based On In-memory Computing

As convolutional neural networks(CNNs)are widely applied to many applications,such as image and video recognition,massive CNN tasks should be efficiently handled in a very short time.Since the limitation of traditional processor architectures such as CPU and GPU,these processors can not efficiently perform the computations of neural networks.Many related research works proposed deep neural network(DNN)accelerators based on near-storage computing.But these Von Neumann-based accelerators are unable to essentially estimate the negative impact caused by the memory wall problem.Therefore,some researchers proposed the in-memory computing-based DNN accelerators to solve the memory wall problem generated by a number of data access.Most existing works designed the in-memory computing-based DNN accelerators with many fixed-size and independent crossbar processing units.The fixed-size and independent crossbar processing units can perform various matrix multiplications which have different-size inputs and outputs by splitting the multiplication to several parts and parallelly computing in crossbar processing units.Since the crossbar processing units compute in analog signals and analog signals can not be stored,the crossbar processing units need digital-to-analog converters(DAC)and analog-to-digital converters(ADC)to convert signals,which will generate a mass of time and energy consumption.Focus on the problems referred above,I propose an in-memory computing-based CNN accelerator.The main works in the thesis are as follows:Analyzing the structural characteristics of classic convolutional neural networks,the research status of deep neural network accelerators and the shortcomings of deep neural network accelerators based on memory computing.The ReRAM-based Convolutional Neural Network Accelerator Architecture(RFSM)is proposed,and the key technologies contained in RFSM are explained in detail.With using the relationship between inputs and outputs between adjacent layers,receptive-field based convolution srategy and dynamic and configurable combinations for crossbar are proposed for improving the processing efficiency of inference tasks.Receptive-field based convolution srategy guides the crossbar array sets to dynamically connect to perform multi-convolution calculations without signal conversion operations inside the crossbar array sets.This paper also proposes a single-image inference task M division technology,the inference tasks after the division of this technology can be processed independently of each other in parallel.I use CACTI 6.5 to model energy and area for all buffers and on-chip interconnects.The ReRAM-based crossbar array energy and area model are based on research.The power and area models of ADCs and the DACs are cited from the research.The evaluation result shows that,compared to existing works(ISAAC),RFSM gains up to 6.7x higher speedup and 7.1x lower energy consumption.