Sparse GRU neural network acceleration implementation method and device

Abstract

The invention provides a sparse GRU neural network acceleration implementation method and device, and the method comprises the following steps: S1, training a GRU neural network model through employing a CPU or GPU, and carrying out the pruning and quantification of the trained model parameters; storing the sparse parameter matrix by using a triple method, and quantifying the input of the model; S2, transmitting the triple for storing the model parameters and the quantized input to an external memory of an FPGA by using a buffer; and S3, realizing calculation of the sparse GRU neural network in the FPGA, and transmitting a final result to the external memory. According to the sparse GRU neural network acceleration implementation method and device, the calculation efficiency can be improved, the input transmission time and the data transmission frequency are reduced, and then power consumption and time delay are reduced.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a method and device for realizing acceleration of a sparse GRU neural network. Background technique [0002] In recent years, the rise of deep learning has continuously promoted the development of artificial intelligence. As an important tool for deep learning, the deep neural network guarantees the fitting ability of the model by increasing the number of layers of the model and increasing the amount of training data, which brings about an explosive growth of model weight parameters and calculation volume. In order to improve the performance of neural networks, heterogeneous computing solutions are often used. At present, most researchers use GPUs to accelerate deep neural networks. Graphics processing units (GPUs) are used for computationally intensive tasks and have the characteristics of high bandwidth and high parallelism, but they have high power consumption. The p...

AI Technical Summary

Problems solved by technology

Usually researchers use GPU to accelerate the GRU neural network, but due to the timing-dependent nature of the GRU neural network, it is difficult to achieve a high degree of parallelism

Most of the existing FPGA neural network accelerators are designed for convolutional neural networks or fully connected neural networks, and have less support for GRU neural networks

At the same time, most deep neural networks have the characteristics of many parameters and a large amount of calculation. How to use FPGA with limited resources to accelerate is still a difficult point.

Images