Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Hardware acceleration method for long short-term memory neural network and computing system

A long-short-term memory and neural network technology, applied in the field of hardware acceleration methods and computing systems, can solve problems such as large amount of calculation, long calculation time of nonlinear functions, and high demand for computing resources

Active Publication Date: 2022-07-08
저장진셍일렉트로닉스테크놀러지컴퍼니리미티드
View PDF12 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0010] (3) The amount of calculation is large. A long-term short-term memory neural network contains hundreds of long-term short-term memory units. Assuming 100, the internal matrix calculation process is the same as that of the fully connected layer, so doing a long-term short-term memory neural network calculation is equivalent to doing 800 Computing at the sub-full connection layer requires high computing resources
[0011] (4) The calculation time of the non-linear function used by the internal 4 gate circuits is long
However, in this prior art, the main disadvantage of its design is that the fully connected layer is designed separately, and the digital signal processor is only used to calculate the matrix multiply-accumulate operation in the long-term short-term memory layer, and its versatility is not strong

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Hardware acceleration method for long short-term memory neural network and computing system
  • Hardware acceleration method for long short-term memory neural network and computing system
  • Hardware acceleration method for long short-term memory neural network and computing system

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0035] In view of the demand for hardware acceleration of long-term and short-term memory neural networks, the publication proposes a method for hardware acceleration of long-term and short-term memory neural networks and a computing system for implementing the method. The computing system can be a computer with processing circuits, memory, and related software and hardware implementation System, embedded system or single-chip system, for the problem of large amount of parameters, the floating-point number can be converted into 8-bit fixed-point number for calculation to reduce the amount of parameters; and for the problem of large amount of calculation, both area and speed are considered; And it can make full use of the existing fully connected layer unit of the hard core accelerator to realize the matrix multiply-accumulate operation of the long short-term memory neural network, reduce the redundant design, and improve the design versatility. In particular, for nonlinear func...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a hardware acceleration method for a long short-term memory neural network and a computing system for executing the method, the method comprises the following steps: firstly, quantizing parameters of the long short-term memory neural network, performing floating-point computation on floating-point data input into the long short-term memory neural network to obtain output floating-point data, and performing quantization according to the output floating-point data; and performing fixed-point calculation in the long short-term memory neural network to output fixed-point data. Then full-connection calculation is executed, specifically, weights in a full-connection layer in any long short-term memory unit in the long short-term memory neural network are rearranged, similar items are combined to execute less-order full-connection layer calculation, and then a specific nonlinear function is calculated by using a simplified instruction vector instruction set; after the steps are repeated, the matrix multiply-accumulate operation of the long-short-term memory neural network can be completed.

Description

technical field [0001] This proposal mainly involves the hardware design of neural network, especially refers to a hardware acceleration method and computing system for implementing long short-term memory neural network by means of neural network parameter quantization and using reduced instruction item quantity instruction set. Background technique [0002] Long-Short-Term Memory (LSTM) neural network is a recurrent neural network (Recurrent Neural Network, RNN). Long-short-term memory neural network introduces long-term and short-term memory unit (LSTMunit), through various gates (gate) Control long-term and short-term memory units, regulate the input, output, forget and hidden states of long-term and short-term memory units, and improve the shortcomings of recurrent neural networks in long-term memory, that is, to solve the learned recurrent neural network. The long-term dependency problem that exists in the network. For example, the gate signal is used to avoid gradient...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
IPC IPC(8): G06N3/063G06N3/04G06N3/08
CPCG06N3/063G06N3/049G06N3/08G06N3/045Y02D10/00
Inventor 周志远陆金刚沈强方伟
Owner 저장진셍일렉트로닉스테크놀러지컴퍼니리미티드
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products