Configurable convolution accelerator applied to convolutional neural network

Abstract

The invention discloses a configurable convolution accelerator applied to a convolutional neural network, which comprises a transmission clock domain module and a main clock domain module, and is characterized in that the main clock domain module comprises a computing unit array, a global cache and an on-chip network; the global cache is used for storing the a first intermediate result, the a second intermediate result, the an input feature map and the convolution kernel weight, and transmitting the output feature map; the on-chip network is used for controlling multiplexing of the structure of the computing unit array, the convolution kernel and the input feature map, and controlling the position and direction of data transmission; and the calculation module is used for performing convolution calculation on the acquired input feature map, the convolution kernel weight and the first intermediate result to acquire a second intermediate result and an output feature map. According to theconfigurable accelerator, the structure of the computing unit array is controlled through the network-on-chip, multiplexing of the computing unit on the input feature map and the convolution kernel and accumulation of the second intermediate result are achieved, off-chip storage access of the input feature map and convolution kernel data is remarkably reduced, and then power consumption of the configurable accelerator is reduced.

Description

technical field [0001] The invention belongs to the hardware field of convolution layer operation acceleration, and more specifically relates to a configurable convolution accelerator applied to a convolution neural network. Background technique [0002] Convolutional Neural Networks (CNN) are currently widely used in image recognition, including autonomous driving, drone navigation, and robot vision. Although CNN can provide the highest accuracy in many image recognition tasks, it will greatly increase the computational complexity. Therefore, techniques that can effectively improve the energy efficiency and throughput of processing CNNs without reducing application accuracy or increasing hardware costs are crucial for the widespread deployment of CNNs in artificial intelligence (AI) systems. [0003] In convolutional neural network computing tasks, there are a large number of convolution operations that can be performed in parallel. If a traditional general-purpose process...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the defects of the prior art, the purpose of the present invention is to provide a configurable convolution accelerator applied to the convolutional neural network, aiming at solving the problem of the low data utilization rate in the existing convolution operation which leads to the frequent failure of the convolution acceleration operation unit. External storage accesses data, and then the problem of high power consumption of CNN hardware accelerator

Images