Feedback about CUDA vs OpenCL (GPGPU HPC)

[jchd 1,514]

Hi folks/geeks,

My question is only loosely related to C/C++ and other common topics in this area but anyway there is a tiny but > 0 chance that someone could give valuable unbiaised advice. (Specialized forums tend to get highly religious about the question so opinions there are to be taken with at least 5 grains of salt!)

Does anyone have sound and practical experience of High Performance Computing (HPC) using CUDA (Nvidia-based) versus OpenCL (AMD/ATI-based)?

From what I've gathered over weeks/months, CUDA is easier to learn and code, way more common, enjoys a huge range of available libraries for dedicated domains (which I've no use for).

OTOH OpenCL, while having a steepper learning curve and a reputation of being more involved to code for, is an open standard which is likely to gain more and more acceptance over time.

Also paper-only comparison of competitive ATI vs Nvidia hardware capabilities are a bit fuzzy and mostly game-biaised, even more with the unkown of how well the long-awaited Nvidia "Kepler" architecture will do. (I can wait until Q2 for Kepler series launch).

I'm not much interested in double-precision FP operations since the massive work I'm going to feed to GPUs will be 99.999% integral (essentially linear algebra). I won't have use for any ready-to-use library so the advantage of offering a wide applicative library support doesn't mean much to me.

Of course I'd like to get the maximum possible bang for the buck but I'd as well avoid investing into unduly hard to program hardware.

In the light of this use case, who wins CUDA vs. OpenCL?

The software part will rely on Mathematica v8 (of which I'm a big big fan) to distribute the workload to C threads running on as many GPUs shadder units as possible and gather/process partial results. Version 8 of Mathematica offers incredibly nice support for both CUDA and OpenCL, so the choice is wide open at this level.

TIA for opinions.