FPGA-Based Reconfigurable CNN Accelerator Design

With the continuous development of research in the field of artificial intelligence,convolutional neural networks have become more and more prominent in the fields of artificial intelligence computer image algorithm processing,text information extraction,and voice recognition.As a result,many scholars have devoted themselves to the research of convolutional neural networks,deriving many more complex and deeper layer network structures,with the consequent huge computational workload.Processors with von Neumann structures are often less efficient than dedicated hardware gas pedals when dealing with function-specific accelerated systems due to their complex control operations.However,the development cost and cycle time of dedicated hardware gas pedals are no longer acceptable in the face of today’s complex convolutional neural network structures.Therefore,the characteristics of FPGAs make them the optimal choice for developing gas pedals with high reconfigurability,adjustable parallelism,high performance,and low power consumption.In this paper,we study the design of a reconfigurable modular CNN gas pedal system architecture and implement the application on FPGA platform.The architecture can flexibly adapt to different neural network structures and achieve efficient computational performance through hardware optimization and parallel computing.Firstly,the top-down design of the general system architecture improves the data transfer efficiency and reduces the delay caused by the data access;secondly,the hardware resources are allocated according to the function and connection mode of each layer module,and a large amount of DSP resources are allocated to the convolutional layer module,which has the largest computation volume,to meet the high parallelism of convolutional operations;then,the modules and sub-modules of each layer are designed,and the upper layer module is set to The number of instantiated submodules is used to adjust the parallelism of multiple dimensions.Finally,the modules are implemented in C and then ported to the FPGA platform using the hardware description language.The Verilog language is used for hardware description and module design,and Vivado is used for simulation and synthesis implementation.The performance of the gas pedal was experimentally evaluated and comparatively analyzed through the analysis of experimental results.The reconfigurable CNN gas pedal designed in this paper has a performance of 102.4 GFLOPS,an acceleration ratio of 7.25 times that of CPU,and an energy efficiency ratio of 9 times that of GPU,which is superior to other studied FPGA gas pedals in terms of performance,power consumption and energy efficiency ratio,proving that the The efficiency and feasibility of the proposed gas pedal are demonstrated.The reconfigurable CNN gas pedal design research results in this paper provide a feasible acceleration solution for convolutional neural networks in lightweight and high-precision practical application scenarios,and have certain practical application value.