Hi there fellows and nVidia Friends,
in good old traidition of, and with, those individuals who dared so much as to speak out the forbidden words in this Cuda-sacred forum: << Sorry for asking, but what’s driving nVidia to obviously keep -sorry again- filibustering a truely supportive OpenCL pacakge ??? >>
Good experience (non nvidia):
My corporate laptop runs an Intel HD4000 Ivy-Bridge lower end device, yet it’s at least OpenCL 1.2 compliant and munches perfectly smooth even on that niche-like PyOpenCL package without having to rebuild the native part of that widely used and pretty matured library against BOOST and Intels SDK sources.
Poor experience (nvidia):
For like a decade it’s been that every two years I used to purchase a new nVidia (!) graphics card for my home computer - so far the “pick algo” has been driven and triggerd mainly by gaming stimuli :). Yes, I belong to those who have been conquered by nVidia by the time they acquired 3DFX :) - and since was loyal !!!
But currently Im seriously thinking of getting to embark to new lands and sail in with AMD’s harbor maybe.
Right now my home PC sports a somewhat outdated GF116 powered GTX550 Ti featuring some neat 3GB DDR3 RAM (but the 1G DDR5 mentioned below):
Absolutely nice, smooth, fast enough for what gaming I do every now and then and supposedly sufficiently equipped (192 cores and 3Gigs RAM, c’mon) to get a neat OpenCL snippet kicked off and compute what I usually develop primarily on my corporate laptop (again, intel platform) and which finally is going to end up in an embedded device anyway, thus both, Intel and nVidia platforms were makeshifts only, albeit useful ones.
Due to portability, CUDA is a no go, for the targetted embedded platform doesnt feature it.
Now, everything were fine if it wasnt for nVidia’s somehow hard-to-reconstruct OpenCL support-notsupport policy.
First of all, tech specs and OpenCL as a search term as such are incredibly hard to find on whatever nVidia website or branch thereof. Secondly, if the term is found, the link leads either nowhere or into a massive CUDA document filled with an abundance of anything but OpenCL related chapters / paragraphs. Totally misleading, and, since seemingly the default case: intention ? Valued users have to plow through that doc and usually end up seeking “OpenCL” using CTRL-F. In the end the info surfaced usually lacks detail or is devoid of key data one has to have to localize where standing at.
According to public sites (like two above), the GF116 GTX550TI supports OpenCL1.1 and yes, in fact, it does so, as I managed to get some code executing on it even with PyOpenCL2014.1 (this time i had to rebuild the native part against BOOST and nVidias sources, using MinGW).
Strange enough, though, since the comprehensive Khronos conformant products list doesnt mention the GF116 or family members at all. Just adds to the messy impression of that OpenCL situation I gained with nVidia:
Now, 1.1 and 1.2 do have substantial differences (Barriers changed, MEM flags added, some functions renamed and merged and signatures changed) and it’d be useless mapping what will end up with rev 1.2 temporarily to 1.1 just for the sake of dancing the way one is forced to in order to utilize a graphics card for dev-platform only purposes cuz its vendor decided to stick a label to it but instead of really supporting what the label says --> OpenCL compliant <-- only provide grand-mother’s revision dating back a millenium.
---->>>> Does somebody know, or believe to know, whether nVidia is going to release an OpenCL1.2 conformant update any time soon, moreover, whether this -potential update- will support the GF116 GTX550Ti, too ?
(I wouldnt care of, and I guess nobody would dare to ask for, an OpenCL-devoted branch on nVidias website which compared to the cozy, neat corner those CUDA enthusiasts are supplied with in terms of tech briefs, white papers and app notes. All we ask for is: Sell us a Ferrari but keep from determining which race league costomer’s gotta join - if it’s CUDA or the league of OpenCL. If you dont really wanna support one - really do say so.)
If nobody knows nothing, or in case of a negative statement I’d guessed I were required to search for AMD cards to purchase and ditch my veteran GTX550 - what a shame.
Anyways, thanks buddies, live long and prosper \// !