You've got your GPU program that needs to be compiled to actual GPU code at startup as well. That takes a bit. The actual file transfer is comparable to the GP

You've got your GPU program that needs to be compiled to actual GPU code at startup as well. That takes a bit. The actual file transfer is comparable to the GPU doing a memcpy. (You could also use pinned host memory if the GPU program is just doing a one-off read.) GPU processing for simple operations really shines if your data is already on the GPU and your code is hot. Otherwise, it's more useful if the process is significantly complex enough to make the overhead of a cold startup insignificant.