Computing unit, array, module, hardware system and implementation method

Abstract

The invention discloses a computing unit, an array, a module, a hardware system and an implementation method, and belongs to the field of artificial intelligence algorithm hardware acceleration. Aiming at the problems of huge data and long calculation time of a sparse convolutional neural network algorithm in the prior art, an invalid data removal mechanism is designed in a calculation unit, invalid weights or input image data can be removed, the calculation time is reduced, and the power consumption caused by multiplication and accumulation calculation is reduced; a multi-channel sub-computing unit is designed, a multiplexing accumulation channel mechanism is adopted to complete convolution operation, and resource consumption is reduced; under the condition that invalid data is removed, asupply number rotation mechanism is further designed, and sufficient supply number of the calculation unit can be kept; the method is low in power consumption, small in area, high in throughput rateand high in recognition speed, is suitable for application of mobile terminals, such as smart home and smart city, and can efficiently complete license plate recognition, face recognition and the like.

Description

technical field [0001] The invention relates to the field of artificial intelligence algorithm hardware acceleration, in particular to a computing unit, an array, a module, a hardware system and an implementation method. Background technique [0002] Convolutional Neural Network (CNN) is a feedforward neural network that has a wide range of applications in the field of artificial intelligence, including image recognition, big data processing, and natural language processing. In order to improve the accuracy of the algorithm, the model structure of the convolutional neural network is becoming more and more complex and the depth is increasing. The resulting large model parameters and long calculation time hinder the deployment of the algorithm in terminals, such as smart homes and smart transportation. and other IoT applications. These issues have led to intensive research on the algorithm and hardware design of convolutional neural networks in pursuit of low power consumptio...

AI Technical Summary

Problems solved by technology

Although this invention also improves the overall efficiency from the perspective of hardware architecture, it does not process invalid data, resulting in high power consumption and insufficient optimization of calculation time

Images