Research And Implementation Of CNN-Oriented High Energy Efficiency SRAM Computing Array

Convolutional Neural Networks(CNN)has achieved excellent classification results in the field of visual perception,which has completely changed the visual framework of traditional computers.However,the problems of memory access and computational power consumption seriously limit the development speed of CNN.Therefore,how to break through the limitation of memory and achieve efficient convolution computing is of great significance.One of the ideas is to use binary networks to reduce the amount of data,thereby reducing power consumption.In addition,computing in memory(CIM)breaks through the traditional Von Neumann architecture and shows its unique advantage in the field of energy efficient computing.In order to solve the problem,a CNN-oriented high energy efficiency computing array based on Static Random Access Memory(SRAM)is proposed,which both has convolutional computing mode and SRAM memory mode and composes a CIM architecture.Besides,convolutional computing mode is realized by pulse modulation in time domain.The design idea of SRAM computing unit is given at first,then it could be improved from two aspects of storage structure and calculation structure.Furthermore,a pulse accumulator quantizer based on capacitor discharge is put forward to optimize quantization circuits.Finally,the power consumption of SRAM array is specially optimized according to the network structure.The proposed scheme is implemented based on TSMC28nmHPC+ process,and has been submitted for tape-out.The size of SRAM array is 896*224 bits.The simulation result shows that the write yield in SRAM mode reaches 100% under TT corner,25℃,0.9V and 500 MHz,which meets the design requirements.At the same time,in the computing mode,the energy efficiency achieves 39.03 Tops/W with AlexNet as benchmark,which is 17.74 x and 1.96 x of other digital designs.Compared with CIM architectures,the scheme has the advantage of network scale.