Originally posted by: erman_amd nou,
I already read that before I ask in this forum. If not, why did I ask?
"please provide official documentation for the GPU architecture, not only the OpenCL programming". I think there several architectures used in AMD GPU.
No wonder people choose N***** than AMD, in research.
I'd argue they picked nvidia not because nvidia has more information, as they do not, but because CUDA worked years before AMD had anything that you could download that remotely worked anyhow.
In 2007 i phoned dozens of times to AMD officials and the only thing they wanted to do was sell me something, not interested in delivering support at all, though that was the official guy delivering AMD support for GPU hardware.
Maybe i should claim the 460 euro i paid for that card on AMD, as there wasn't a CTM released unlike announced 
You know, gpu's are pathetic documented. What you can do however is dig up some NSA type guys who lurk around. Usually doing something with nvidia, what's the case with nvidia usually works for AMD in a similar manner and those dudes happen to 'by accident' know that.
Note Nvidia isn't very clever either. For worlds biggest simulator software. No rival on the planet, it's getting used from aviation industry to car industry to military, i tried talk about setting up a pilot project.
at that time
a) opencl wasn't there (which would have been a requirement for the project, as they don't want to rely upon just 1 manufacturer for their software to run at)
b) nvidia tesla manager first wanted me to first sign for a considerable amount of gpu's getting sold for this project; though we didn't talk numbers i had impression he referred to thousands. Realize a guy, me, who tries to setup a pilot project that would sell 'em a couple of millions of tesla's when it would've succeeded and shown favourably fast over existing hardware, to first sign for garantueed sales, for a PILOT project.
But of course biggest problem was only support *during* the pilot project could get garantueed; that would be completely unacceptable for these guys. They don't build worlds biggest clusters just for fun; years of support needed of course long after such pilot project terminates.
In this sense the GPU managers are total idiots with respect to gpgpu, they have *no clue* how things work in the world of computation. You must lay down on your knees to get new hardware platforms accepted in pilot forms, not lay down on your knees for a small piece of information with respect to how things work on those gpu's.
I wouldn't even dare approach them with Radeon 6990 cards, as the obvious question would be: do they already have a driver and get supported by opencl (so the full 3072 PE's, not just 1536 of 'em)?
I
Jawed, thanks for the answer.
diepchess, thanks for sharing your thoughts.
It's not my intention to compare whether AMD or NVIDIA is better.
I choose AMD platform because not so many people use it out there (at least now and based on my paper survey). Many research paper use NVIDIA CUDA, so I want to try different one. Also since high school my PC use AMD CPU
