A user-defined floating-point number and a calculation method and a hardware structure thereof

Abstract

The invention provides a user-defined floating-point number with a shared index. A 32-bit floating-point based neural network model to 8-bit, which greatly reduces the size of the model while reducingthe computational complexity. In user-defined floating-point operations, integer multiplication and addition are used. Compared with 32-bit floating point multiplication, the multiplication of user-defined floating point numbers saves 17 times of the energy consumption, 30 times of the chip area; the addition operation saves 28 times of the energy consumption and 116 times of the chip area.; in the network parameters of the full connection layer, the memory bandwidth required by the data transmitted to the off-chip memory is reduced by four times. The user-defined floating-point number is helpful for maintaining parameters over four times in the on-chip buffer. In the hardware implementation, combined with the 8-bit user-defined floating-point number and multiplier package structure, theoperation speed and throughput of the entire neural network are increased on the operation unit CU..

Description

technical field [0001] The present invention relates to the technical field of convolutional neural network, specifically, relate to a kind of self-defined floating-point number and its calculation method and hardware structure, especially design relates to a kind of based on self-defined floating-point and shared exponent, in FPGA convolutional neural network The multiplier package structure. Background technique [0002] In recent research on convolutional neural networks (CNNs), most models pay little attention to their complexity in order to maximize accuracy. For example, in modern deep CNN models, such as AlexNet, GoogleNet, and ResNet, all require millions of parameters and billions of arithmetic operations, the extremely high computational complexity and massive resource consumption hinder the development of embedded devices. accomplish. [0003] Reducing the representation precision of data is a common way to speed up training and reduce memory bandwidth. In term...

AI Technical Summary

Problems solved by technology

Dynamic fixed-point numbers can simplify calculations, but due to the limited range of decimal point positions, they are usually only suitable for the reasoning process of CNNs, and cannot be adapted to the training of large-scale CNNs

In the latest research on floating point numbers, the mini floating point representation method (9-bit) is used, but the results only give the top 5 accuracy based on Cifar-10

[0004] In the practice of hardware acceleration, GPU is widely used in training and classification tasks, however the power consumption of GPU (100W) is too high for custom data representation methods

Images