A Deep FM Acceleration Algorithm based on Baidu Cloud’s Heterogeneous Multicore Platform

 A deep FM acceleration algorithm was designed based on Baidu Cloud’s heterogeneous multicore platform that contains both Intel’s E5-2650 v4 CPU and Xilinx’s XCKU115 FPGA.

A brief introduction of our work is as follows.

1)  Firstly, the model was trained using Baidu’s open-source-deep-learning framework called PaddlePaddle.

2)  In order to better understand how the model works, the whole system were rebuilt using C++, preprocessing procedure, parameter extraction and model reconstruction included.

3)  What’s more, computing tasks were split according to the reconstructed system. Only fully connected layers were accelerated by implementing them using Baidu’s Polaris API that helps accelerate programs with FPGA.

4)  DMA bandwidth was utilized as full as possible, while the result, Polaris API used, was assured to be correct.

5)  Last but not least, the performance of our algorithm is about 4 times better than the original version that runs at a PC equipped with Intel’s i5-3470 CPU, with the time consumed by DMA not taken into account. The performance of our algorithm is more than 2 times of the original version, DMA time included.

62

Figure 1. System Overview

63

Figure 2. Algorithm Framework

发布时间:2019-08-21 22:48
浏览量:0
点赞
收藏
Home    Catalog    2018 OpenHW    A Deep FM Acceleration Algorithm based on Baidu Cloud’s Heterogeneous Multicore Platform