Sparse accelerator applied to on-chip training

Abstract

According to the sparse accelerator applied to on-chip training provided by the embodiment of the invention, a plurality of input numerical values in the input numerical value buffer module, a plurality of reference numerical values in the reference numerical value buffer module and a mask in the mask buffer module are dynamically adjusted in different acceleration stages; the method comprises the following steps of: performing primary screening on invalid operations by a coarse-grained unit in a coarse-grained manner, and further screening the invalid operations by a fine-grained unit included in each processing unit in a processing module in a fine-grained manner, thereby eliminating all invalid operations in three training stages of on-chip training; in addition, the plurality of computing cores can perform parallel acceleration processing, and the plurality of processing units included in the processing modules in the plurality of computing cores can also perform parallel acceleration processing, so that the hardware utilization rate of the sparse accelerator is further improved. Therefore, the sparse accelerator applied to on-chip training can efficiently and accurately eliminate all invalid operations in three stages of on-chip training.

Description

technical field [0001] This application relates to the field of computer and electronic information technology, in particular to a sparse accelerator applied to on-chip training. Background technique [0002] In recent years, Convolutional Neural Networks (CNN) have performed well in the fields of computer vision, speech recognition, and natural language processing. In order to improve the accuracy of recognition, it is necessary to train the CNN model on-chip. Fine-tune the CNN model with the user data on the network, and then improve the training accuracy of the CNN model when it is used. [0003] The on-chip training process includes three stages, namely the forward propagation (FP) stage, the backward propagation (BP) stage and the weight gradient calculation (weight gradient, WG) stage. In the FP stage, the previous layer The activation value of the current layer is used as the input activation value of the current layer, combined with the convolution kernel weight cor...

AI Technical Summary

Problems solved by technology

[0006] This application provides a sparse accelerator applied to on-chip training, which is used to solve the technical problem that after the acceleration of the existing accelerator, all invalid operations in the on-chip training cannot be eliminated, and the most efficient implementation of on-chip training on the terminal device cannot be guaranteed.

Images