A Highly Compatible Programmable Neural Network Acceleration Array

Abstract

The invention belongs to the technical field of integrated circuits, specifically a high-compatibility programmable neural network acceleration array. The array employs a reconfigurable architecture,and comprises one central controller, one feature vector transmitter, and a plurality of neural network calculation unit pieces. The neural network calculation unit pieces comprise a programmable multiplication and addition unit, a programmable activation unit and a unit piece controller, and the acceleration array carries out the communication between any unit pieces through a programmable communication route. The acceleration array can be compatible with various types of neural network algorithms, does not lose the high energy efficiency, and is suitable for various types of deep learning intelligent systems.

Description

technical field [0001] The invention belongs to the technical field of integrated circuits, and in particular relates to a highly compatible programmable neural network acceleration array. Background technique [0002] Nowadays, the development of custom deep learning acceleration chips on mobile devices is becoming more and more popular. The challenge is that the performance of chips is limited by the type of deep learning network, such as CNN (convolutional neural network), RNN (recurrent neural network), in order to design high Energy-efficient custom deep learning accelerator chips are often optimized for certain networks, with high performance when using these networks, and poor performance under other networks. However, due to the recent rapid development in the field of deep learning, an improved version of CNN or RNN network may appear in the future, and even other new deep learning neural network algorithms will appear, so the existing dedicated deep learning accele...

AI Technical Summary

Problems solved by technology

[0002] Nowadays, the development of custom deep learning acceleration chips on mobile devices is becoming more and more popular. The challenge is that the performance of chips is limited by the type of deep learning network, such as CNN (convolutional neural network), RNN (recurrent neural network), in order to design high Energy-efficient custom deep learning accelerator chips, which are often optimized for certain networks, perform very well when using these networks, and perform poorly on other networks

However, due to the recent rapid development in the field of deep learning, an improved version of CNN or RNN network may appear in the future, and even other new deep learning neural network algorithms will appear, so the existing dedicated deep learning acceleration chips will not be able to achieve the desired results. The required performance requirements fundamentally limit the development of deep learning intelligence

Images