Data compression method, data compression system and operation method of deep learning acceleration chip

Abstract

The invention discloses a data compression method, a data compression system and an operation method of a deep learning acceleration chip. The data compression method includes the following steps. And obtaining a filter coefficient tensor matrix of the deep learning model. Performing a matrix decomposition program according to the filter coefficient tensor matrix to obtain at least one sparse tensor matrix and at least one conversion matrix; the product of the conversion matrix and the filter coefficient tensor matrix is a sparse tensor matrix. The conversion matrix is an orthogonal regular matrix. And compressing the sparse tensor matrix. And storing the sparse tensor matrix and the conversion matrix, or storing the sparse tensor matrix and the reduction matrix in a memory. And the deep learning acceleration chip performs operation by using the sparse tensor matrix to obtain a convolution operation result. And the deep learning acceleration chip restores the convolution operation result by using the restoration matrix.

Description

technical field [0001] The invention relates to a data compression method, a data compression system and an operation method of a deep learning acceleration chip. Background technique [0002] With the development of deep learning technology, a deep learning acceleration chip has been developed. Through the deep learning acceleration chip, complex convolution operations can be directly calculated in hardware to speed up the operation. [0003] The deep learning acceleration chip is paired with a large-capacity memory to exchange temporary data through the system bus. Ideally, data movement and operation are parallelized at the same time. However, in practice, due to shared busbars or physical limitations, most of the time is occupied by data movement, resulting in lower performance than expected, and external memory accesses cause major power consumption. Amdahl's law also shows that the use of parallelism to improve the performance still has its upper limit. From the comp...

AI Technical Summary

Problems solved by technology

How to break through this slope limit has become a major technical bottleneck for deep learning acceleration chips

Images