Addition network-oriented near-storage neural network accelerator and acceleration method thereof

Abstract

The invention discloses an addition network-oriented near-storage neural network accelerator and an acceleration method of the near-storage neural network accelerator. The neural network accelerator comprises an instruction generation unit, a calculation unit group and a post-processing module. The accelerator compares the low-bit weight and activation in the instruction generation unit to generate an instruction, the instruction is transmitted to a specific calculation unit in the calculation unit group to guide calculation of the full-precision weight or activation stored in the specific calculation unit, output is transmitted to the post-processing unit according to the instruction to be processed, and a final result is generated. According to the method, symbol extraction and full-precision data calculation are separated, a reconfigurable and extensible near-storage calculation unit is designed, different neural network reasoning tasks can be conveniently adapted. Meanwhile, data compression and grouping are carried out on an addition network model, only non-sparse full-precision data are reserved, and the model size and the calculation frequency are reduced.

Description

technical field [0001] The invention belongs to the technical field of deep neural network acceleration computing, and in particular relates to an additive network-oriented near-storage neural network accelerator and an acceleration method thereof. Background technique [0002] Convolution Neural Network (CNN) is a machine learning algorithm that has achieved good results and has been widely used in computer vision, speech recognition, natural language processing and other fields. But the convolutional part of the CNN algorithm is computationally intensive and takes up huge storage space, so when we deploy convolutional neural networks on end devices and embedded devices, we usually use application-specific integrated circuits (ASICs) to accelerate them to Achieve low power consumption and high performance features. [0003] The additive neural network (AdderNet) is a kind of convolutional neural network, which uses the sum of absolute difference (Sum of Absolute Difference...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a near-memory neural network accelerator and its acceleration method oriented to the addition network, so as to solve the technical problems of large power consumption, low reasoning speed and low efficiency in the existing reasoning calculation process

Images