Thanks for your answer.
I see. But in my case, when we need to allocate a very large memory in GPU, that exceeds the size of one card, maybe the size of 2 cards is enough. I cannot know any solutions for this.
My alogorithm is like this, we have a fixed and very large memory for the result, called structure A. Every time, one threads runs, it gets some information from A, computes something and updates the A.
Could u have any recommendations for the problem like this.
Can we do something like this, we allocate the memory for A in two cards 1 and 2. Every time, one threads run, depends on the information it has, it knows (in fact it must know) the information it needs is in card 1 or card 2. And it takes what its wants, computes and updates A.