Graphics Processor Clusters for High Speed Backpropagation
Sensors and Electromagnetic Applications Laboratory, Georgia Tech Research Institute
Fifteenth Annual Workshop High Performance Embedded Computing (HPEC 2011), 2011
This paper describes the use of GPU clusters to accelerate backpropagation for Synthetic Aperture Sonar (SAS) systems. We extended a GPU-based implementation of backpropagation to support clusters of GPU-enhanced nodes. The GPU accelerated implementation formed a 3,936 x 3,936 SAS image from 60s of sonar data in under 12s using a single GTX480, and under 3s on a small cluster with 8 Tesla C2050s. Our presentation will discuss the details of our implementation, describe our optimizations, and show performance for various image sizes and execution platforms.
January 14, 2012 by hgpu