Convolutional Neural Network Hardware Accelerator System Based on Convolution Kernel Splitting and Its Computing Method

Abstract

The invention discloses a convolutional neural network hardware accelerator system based on convolution kernel splitting and a calculation method. The system comprises a zero filling module, a controlmodule, a convolution kernel and data splitting module, a convolution kernel weight cache module, a data cache module, an on-chip address index module, a core calculation module and an intermediate result cache module. The zero filling module is used for carrying out zero filling processing on the convolution kernel weight and the picture data; the control module is used for controlling the related modules to operate; the convolution kernel and data splitting module is used for generating a splitting control signal; the convolution kernel weight caching module and the data caching module areused for storing zero-filled convolution kernel weight and picture data; the on-chip address index module is used for generating an address index; the core calculation module is used for calculating data; and the intermediate result caching module is used for storing an intermediate calculation result. According to the system, the operation parallelism can be improved, the hardware complexity is reduced, and therefore the system is suitable for large-scale convolution calculation.

Description

technical field [0001] The invention relates to the calculation of large-scale neural network, in particular to hardware parallel acceleration calculation of large-scale convolution calculation of neural network. Background technique [0002] The convolutional neural network (CNN) first originated in the 1960s. Two neurobiologists, Hubel and Wiesel, discovered that different cells in the cat's visual cortex would be activated according to the direction of light, and thus established a network of cells. The model of activation and transformation of images laid the foundation for the emergence of convolutional neural networks. By 1980, the Japanese scientist K. Fukushima proposed the concept of a neurocognitive machine, which is considered to be the first implementation prototype of a convolutional neural network. [0003] With the rise of artificial intelligence deep learning in recent years, convolutional neural networks have received more and more attention. On the one ha...

AI Technical Summary

Problems solved by technology

Although convolutional neural networks can theoretically be implemented in large-scale parallelism in hardware, they are limited by bandwidth, computing resources, storage and other resources, especially in the face of large-scale convolution calculations.

Images