[profquail]

You don't need to buy a $300 graphics card to experiment with GPU programming.

The current and previous generation Intel CPUs (Haswell and Ivy Bridge, respectively) have on-die GPUs which support OpenCL: http://software.intel.com/en-us/articles/intel-sdk-for-openc...

AMD's APUs are quite cheap (~$100) CPU+GPU designs similar to those in the upcoming PS4 and XBox One (though the retail APUs are somewhat less powerful). They've been more-or-less designed specifically around the needs of a heterogenous OpenCL application.

Finally, the last several generations of NVidia cards all support both CUDA and OpenCL; the newer cards do support additional features though. You should be able to pick up a low-end, recent-edition Nvidia GPU for roughly $100.

The new g2.2xlarge instances are $0.650/hour, and the existing cg1.4xlarge are $2.100/hour; so it may make sense to experiment on AWS a bit, then buy your own card for long-term use if you decide to spend more time doing GPU programming.

[idupree]

Sadly, Intel's integrated-GPU OpenCL still doesn't support Linux, and only just started supporting OS X in 10.9 Mavericks[1]. Usually Intel's Linux GPU support is great; I don't know why this is different.

(Intel do have a Linux OpenCL implementation for Xeon CPU cores and Xeon Phi coprocessor[2], which doesn't help me much. On-CPU OpenCL is fine but hardly faster than regular CPU code, and Phi coprocessors aren't very common currently.)

[1] http://forums.macrumors.com/showthread.php?t=1620203 [2] http://software.intel.com/en-us/vcsource/tools/opencl