Why bring Hadoop into it? MR is a paradigm for parallel programming, and having it compiled into OpenCL for you is hugely convenient. But the best Hadoop use cases always leverage massive storage capacity: people who do ML on Hadoop are doing it because the data is already in Hadoop, not because it seemed like a nifty thing to do. Do you actually leverage HDFS and the Hadoop scheduling infrastructure to run ParallelX jobs? Or can you safely jettison them and just sell this as: "Write your ML/graph analysis/whatever code in Java, have it execute on a GPU"

Also, what are the limits like for input data sizes? I've done a little OpenCL, but I've never gone past the GPU RAM size.