I'm working on a problem that needs a huge array and the 128MB limit is really to small for me . I was wandering if this limit will be increased in the feature versions, or I can't rely on that.
You can play with environment variables to use more of your GPU's memory.
This will allow you to use 80% of your GPU's memory. I've noticed some speed degredation when setting this too high, but your milage may vary.
It worked for me. Now I can make allocations up to 204MB (from total 819MB). This is on windows 7 with 5970. With GTX 470, I can allocate I think something like 300MB.
I am not sure if the environment variable is documented somewhere. If such is the case, it might be for experimental purpose and cannot be guaranteed to be available in future.
GPU_MAX_HEAP_SIZE is documented here http://developer.amd.com/support/KnowledgeBase/Lists/KnowledgeBase/DispForm.aspx?ID=123
and GPU_MAX_ALLOC_SIZE AFAIK don't work anymore.