Zero-copy memory

Hi guys

I have the next question.

If my cuda app requires 3Gb of memory and my GPU only has 2Gb, can I work with zero-copy memory to run my app? or the quantity of host memory (for zero-copy) never can exceed the GPU memory?


Disclaimer: I do not have experience with zero-copy memory, so this is not an answer, but just speculation :-)

Zero-copy means that a block of host memory is mapped into the device memory space. According to that, I would say that you could not map more than 2GB of host memory to your device memory space.

Perhaps, this is something you could test by yourself. You may wish to modify the nelem variable of the simpleZeroCopy example of the NVIDIA CUDA SDK and definitely tell if the answer to your question is “yes” or “no”.