I'm still baffled by toms pushing OpenCL. Why use a TITAN for PS CS6, After Effects CS6, Maya etc and NOT turn on CUDA? ...
You stated some of your points several times, but then maybe one has to given the intelligence level
of some readers. Still, I agree with you re the general idea about CUDA, but perhaps Chris would
point out that the OpenCL-related test on pp. 13 does use the same Titan with all CPU/mbd combos
and thus the intent is to reveal how each CPU is able to exploit the available OpenCL resources,
though I have to say - given the CPU ordering - it does look as though the OpenCL path is limited by
something that's only single-threaded, otherwise the 3930K would be a lot quicker. So, this test really
only shows what it shows, namely how well each CPU can exploit available OpenCL processing power,
rather than be any kind of OpenCL vs. CUDA test. To me it merely shows that OpenCL is not very
efficient, because of the 3930K data.
What would indeed have been useful and interesting is to run the same test but using After Effects,
exploiting the Titan's CUDA, and ensuring the test platforms has much more minimum RAM, ie. 32GB
(otherwise, testing Adobe apps reveals nothing except that one doesn't have enough RAM). That would
then show how each CPU/mbd como is able to exploit available CUDA resources from the same card,
which as you point out is the Titan's native GPU acceleration mode.
The existing results allow one to infer the AMD CPUs are not very good for exploiting OpenCL, but I
think it might be an IPC 1-thread issue because of the positioning of the 3930K result. I don't know
whether this effect is specific to NV cards though, so it would be better to run an OpenCL test of this
kind using an AMD card, and then run an equivalent test with an NV card for CUDA. Not only would
this show in more detail how each CPU is able to exploit GPU resources, ie. whether they differ in
their efficacy for OpenCL vs. CUDA (ie. would a CUDA test show the 3930K doing better?), it would
also finally give an example data point for how well OpenCL compares to CUDA for acceleration in
the same application and render test, especially AE (though the way NV's newer cards work for
CUDA is such that running the same test with two 580 3GB cards would be shockingly revealing,
ie. faster than a Titan).
I suggest using a 290 or 290X for the OpenCL test, and a 780 or 780 Ti for the CUDA test (Titan
really isn't relevant anymore - these apps simply don't need the 64bit fp option most of the time).