DeepLearning4j Adds Spark GPU Support
I previously blogged about the BID project from AMPLab, Single GPU-Powered Node 4x Faster Than 50-node Spark Cluster, and also about an individual's experiment on making Spark's use of blas be GPU-powered, Spark Gets GPU in the Lab.
But a few months ago, the DeepLearning4j project, which already had GPU support, added support for Spark. In contrast, Project Tungsten over in the effort to improve the Apache Spark distribution is not anticipating adding GPU support until Spark 1.6 at the earliest.
Now, there are no benchmarks yet available for DL4J on a GPU-powered Spark cluster. The reason is that they want to make sure they get good apples-to-apples benchmarks done before publishing them.