Heterogeneous storage and calculation fusion system and method supporting deep neural network reasoning acceleration

Abstract

The invention discloses a heterogeneous storage and calculation fusion system and method supporting deep neural network reasoning acceleration, and the system comprises a host processor which is usedfor controlling and managing the whole heterogeneous storage and calculation fusion system; a nonvolatile memory module which is connected with the host processor and used for neural network processing; a 3D stacked memory module which is connected with the host processor and used for neural network processing; a network module which is connected with the host processor and used for being connected with an external host; a configuration circuit which is connected with the host processor, and is used for receiving a configuration command of the host processor and controlling a voltage generator, and is also used for receiving the configuration command of the host processor and configuring a 3D stacked memory module; and a voltage generator which is respectively connected with the nonvolatile memory module and the configuration circuit and is used for receiving the control command of the configuration circuit, applying external excitation to the nonvolatile memory module and adjusting the conductivity state of the nonvolatile memory module.

Description

technical field [0001] The present invention relates to the field of heterogeneous computing acceleration, and more specifically, to a heterogeneous memory-computing fusion system and method that supports deep neural network reasoning acceleration. Background technique [0002] Deep Neural Networks (DNNs) have been used in a wide range of applications, such as object detection, image classification, speech recognition, action recognition, and scene understanding. Deep neural networks are usually characterized by a large number of model parameters (thus limited by memory capacity) and high throughput (limited memory bandwidth when data is moved), requiring a large number of parallel computing and memory access. Some of the deep neural network accelerators proposed in recent years are based on 3D stacked memory, and some are based on new non-volatile memory, and good research results have been achieved. [0003] At present, the acceleration system for deep neural network comp...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the embodiments of the present invention is to provide a heterogeneous memory-computing fusion system and method that supports deep neural network reasoning acceleration, and comprehensively consider the requirements of DNN and the advantages and disadvantages of 3D stacked memory and memristor, so as to solve the existing problems based on 3D stacked memory. The low energy efficiency of the scheme and the limited storage capacity of the memristor-based scheme. At the same time, the system has optimized the characteristics of deep neural network computing and storage access.

Images