FPGA (Field Programmable Gate Array) platform-oriented recurrent neural network algorithm optimization method

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an FPGA platform-oriented recurrent neural network algorithm optimization method. The method comprises the following steps of: storage optimization: storing intermediate variables in a neural network algorithm by adopting a distributed RAM (Random Access Memory) constructed by an LUT (Loop Under Test) in an FPGA (Field Programmable Gate Array), and storing parameters with storage resource requirements exceeding a threshold value in the neural network algorithm by adopting a BRAM (Block Random Access Memory) in the FPGA; and algorithm optimization: for vector multiplication and addition operation in the forward calculation process of the neural network algorithm, carrying out multi-loop operation on the outer layer of the vector multiplication and addition operation, and carrying out pipeline optimization on the calculation process. According to the invention, the calculation efficiency and the resource utilization rate of the recurrent neural network are improved, and the calculation time delay and power consumption are reduced.

Description

technical field [0001] The present application belongs to the technical field of data processing, and particularly relates to an FPGA platform-oriented cyclic neural network algorithm optimization method. Background technique [0002] Machine learning and deep learning are applied in different scenarios, such as text classification, object recognition, speech recognition, and autonomous driving. Even in some specific application scenarios, the performance of artificial intelligence is better than human judgment, such as Google's alpha go robot trained by reinforcement learning to beat the world champion in the game of Go. Data-driven is an important feature of deep learning, and a large amount of high-quality data can train models with higher accuracy and stronger capabilities. But this also brings challenges to the computing power of computers. At present, most of the training deep learning models use CPU and GPU solutions. Because of the structural characteristics of CPU...

AI Technical Summary

Problems solved by technology

But this also brings challenges to the computing power of the computer

At present, the training of deep learning models mostly adopts CPU and GPU solutions. Because of the structural characteristics of CPU, the efficiency and calculation speed of using CPU to train deep learning models are not as good as those of dedicated graphics processors GPU.

However, GPU-based accelerated training models have the disadvantages of high power consumption and insufficient mobile performance, and are difficult to use in application scenarios that require low power consumption, strong real-time performance, and pursuit of mobile performance.

Images