Hook up a lage amount of GPU's to create a Supercomputer...is it possible?
My goal is to hook up a lage amount of Nvidia GPU's into some sort of Supercomputer to drastically increase the power I get from the CUDA module. I would like to start with a 1 TFLOP+ Nvidia CUDA enabled GPU for about 150 dollars and then buy more of the same type of video card every month and hook them together one by one to eventually create a supercomputer. I like the idea of increasing my computing power a TFLOP at a time and I would like no limit to how many times I can do this. I know I can get a MB with 4 pci-e slots but some sort of inexpensive pci-e hubs I can interlink or detailed(and inexpensive) DIY project to make this happen would be better. Is there a better/cheaper way to do what I am trying to do or can someone give advice on how to make this happen. Thank you.
sorry dear, but that's pretty off-topic
It does concern setting up your environment for CUDA module in an AWESOME way. Power loving users like myself can really benefit from this if a cool answer is supplied. Can I reword the ? maybe to fit site guidelines better. I'd like to be able to run a ton of cuda::detectMultiScales on a 1280x1024 video feed, and the matchTemplate/norm combination could use speeding up so the more the merrier if you get my meaning.