Partition-based Neural Network Weight Parameter Mapping Method

At present, with the continuous development of the application of artificial intelligence, it is no longer the computing speed of the processor that affects the performance of the neural network accelerator system. With the increasing complexity of data types, the energy consumed in executing the data computing process is far less than that consumed by data moving between the storage unit and the computing unit. Therefore, in order to minimize unnecessary data movement and reduce the energy loss caused by data movement, Memory-computing integration technology came into being at the historic moment. Its ultimate goal is to break through the traditional Von Neumann computer architecture and integrate the functions of memory and logic computing. The main computing process of neural network includes convolution layer and pooling layer operation. However, the frequent movement of weight data causes great unnecessary power consumption. Therefore, in-memory computing technology is very suitable for neural network applications, and for neural networks with larger weights, in-memory computing technology can show more advantages. This paper mainly introduces the architecture of neural network accelerator based on in-memory computing and two mapping methods of neural network weight parameters on this architecture.