Convolutional neural networks on a multi-FPGA platform

点击翻译按钮获取中文摘要

Machine learning is the use of computer algorithms that can learn from and make predictions about data. Recent breakthroughs in machine learning algorithms have made machine learning one of the fastest growing applications for high-performance computing. A simple example would be to identify faces as you are about to take a picture using your smart phone. The Convolutional Neural Network (CNN) is a machine learning algorithm capable of high-accuracy object detection in images. CNNs are known to run well on graphic processing units (GPUs), which are parallel processing devices originally designed for doing the graphics on computer displays and now also used for many computing applications. A significant issue with GPUs is that they dissipate significant power. Field-programmable gate arrays (FPGAs) are a programmable chip technology that can also be used to implement computing circuits. An important advantage of FPGAs is that they dissipate much less power than GPUs, but they currently do not perform as well as GPUs on CNN computations. However, the use of several FPGAs may overcome this issue. This proposal will investigate a multi-FPGA approach to CNNs. Evertz, based in Burlington, Ontario, is a company that develops and manufactures high-performance computing systems for the broadcast and film industry. Their systems are based primarily on multiple FPGAs to do video processing and switching video streams from multiple sources to multiple destinations. This proposal will investigate the integration of CNNs into the Evertz platform to enhance their products. An example new feature would be to search through an archive of movies and find the ones where Humphrey Bogart appears.