What's the request of using mapped memory?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

In my case, some data needs to transfer to CPU only once.
Can I use mapped memory to improve the performance instead of using memory copy from GPU to CPU?
If yes, what’s the hardware, SDK and driver I need? It seems CUDA3.0 and GTX285 does not support this feature。

Thanks

2

No, you’ll get better transfer performance just using cudamemcpy. Zero-copy isn’t any faster, and in fact it will be slower since it needs two-way communication (the device needs to send a request to the host, and then the data goes back to the device.)

There may be some work-balance net time benefit exceptions though in that zero-copy may help you DELAY a memory copy… and that may be a computational win if you can hide the memory transfer latency and bandwidth by mixing threads with both computation and memory access. So it’s problem dependent.

3

And in answer to your question, you need toolkit 2.2 or later, and a GPU of device 1.2 or later. You’re fine with the GTX285 and toolkit 3.0… it DOES support zero-copy.