Deep learning accelerator and method for accelerating deep learning operation

Abstract

The present invention provides a deep learning accelerator comprising: a plurality of processing components (PE) divided into a plurality of PE groups for performing a computation of a convolution layer by applying multidimensional weights to input activations to generate output activations; the scheduler is used for scheduling the input data in the input activation and the non-zero weight in themulti-dimensional weight to the processing component according to a control mask; wherein the control mask specifies a position of a zero weight of the multi-dimensional weights, and the plurality ofPE groups generate output data of each output channel in the output activation and share the same control mask specifying the same position of the zero weight. Correspondingly, the invention further provides a method for accelerating the deep learning operation. By adopting the invention, the system performance can be improved.

Description

technical field [0001] The present application generally relates to a deep learning computing architecture, and more particularly, to a deep learning accelerator and a method for accelerating deep learning operations. Background technique [0002] Due to its superior performance, deep learning has been widely used in computer vision, speech recognition, natural language processing, bioinformatics and other fields. Deep learning is a branch of machine learning that uses artificial neural networks that contain multiple hidden layers. A type of artificial neural network called a convolutional neural network (CNN) has been used for deep learning on large datasets such as image data. [0003] However, neural networks are computationally intensive. Most neural network calculations involve multiplication and addition calculations. For example, the core computation of a convolutional neural network (CNN) is convolution, which involves high-order nested loops. For feature extract...

AI Technical Summary

Problems solved by technology

[0003] However, the workload of neural network computation is heavy

Images