Design method of lightweight convolution accelerator based on FPGA

Abstract

The invention discloses a convolution acceleration method based on an FPGA. Firstly, a universal processor is used for analyzing neural network configuration information and weight data, writing the neural network configuration information and the weight data into a RAM, and aiming at external memory access bandwidth limitation and based on design space exploration, determining a cyclic partitioning factor to maximize data reuse, so that the operation performance of the whole network is improved. The FPGA reads configuration information from the RAM to generate an FPGA accelerator, the general processor reads picture information and writes the picture information into the DRAM, and the FPGA accelerator reads picture data from the DRAM, starts calculation and writes a calculation result into the DRAM. According to the accelerator, all the layers can be deployed on an FPGA chip at the same time and run in an assembly line mode, and the operation performance and the data throughput rate are improved.

Description

technical field [0001] The invention relates to the field of convolution acceleration, in particular to a design method based on FPGA lightweight convolution accelerator. Background technique [0002] With the introduction of artificial intelligence (AI) into more and more applications such as consumer electronics, automotive electronics, and industrial control, artificial intelligence is facing unprecedented rapid development, and technologies such as deep learning and neural networks have ushered in a development climax. The larger the neural network, the greater the amount of computation required. Although the traditional VPU can also complete artificial intelligence operations, it is already slightly exhausted due to high power consumption and high latency. Loading artificial intelligence computing power on the VPU can avoid these problems and have higher reliability. Target applications include image capture in in-vehicle systems, assisted driving and automatic parking...

AI Technical Summary

Problems solved by technology

However, the shortcomings of GPU are also very obvious. Although it has high performance, the power consumption is also very huge.

Especially when using the GPU to train the convolutional neural network on the PC side, the power consumption can be as high as hundreds of watts

Due to the shortcomings of large size and high power consumption of GPU itself, its promotion and application on mobile terminals and embedded platforms with small size and low power consumption are limited.

Images