Boosting Java performance using GPGPU [pdf]

[Phithagoras]

http://arxiv.org/abs/1508.06791

[sitkack]

This is a direct comparison of their framework and https://github.com/aparapi/aparapi which you can use now.

The JDK is getting support for GPGPU via Truffle/Graal http://www.oracle.com/technetwork/java/jvmls2913wuerth-20139...

[saganus]

This looks very interesting.

Any way we can download some code to test this out or is this still unpublished code?

[jimc39 → saganus]

The work we describe here is our research prototype, I'm currently working towards making the code base more stable before open sourcing. The plan is to release the code over the coming months.

Feel free to track our progress here - https://github.com/aardvark-systems/tornado

[saganus → jimc39]

Great! thanks

[fbristow]

Here's another project that's providing GPU access from Java code: https://github.com/pcpratts/rootbeer1

[nsxwolf → fbristow]

Too bad it's CUDA and not OpenCL. I wonder what the rationale was. Seems natural to use an open GPGPU framework for a Java application.

[fbristow → nsxwolf]

The author explains why here (3 years ago): https://github.com/pcpratts/rootbeer1/issues/2

Not sure that his explanation is accurate, though.

[RussianCow → nsxwolf]

Probably just because Nvidia cards dominate the market for most GPGPU use cases.

[sp332 → RussianCow]

OpenCL is cross-platform, so it would support nvidia, ATI, Intel, CPUs, etc.

[dkersten → sp332]

AFAIK only OpenCL 1.1 (or 1.2?) is available on nvidia though, not OpenCL 2.0 or 2.1. So, unfortunately, if you want to support nvidia (eg AWS GPU instances), then OpenCL may not be the best choice :-(

[sp332 → dkersten]

Oh, I didn't realize it was that bad. Looks like they barely implemented 1.2 and haven't announced support for 2.x at all. https://en.wikipedia.org/wiki/OpenCL#Timeline_of_vendor_impl...

[sitkack → sp332]

If someone made a decent OpenCL 2.X to PTX compiler I think they would do pretty well.