Asynchronous memory copy from Host to Device

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

My application requires constantly streaming data from a live camera feed. I’m therefore really interested in the asynchronous memory copy from host to device.

Are there any nVidia graphics cards currently out that support this capability? I know that cudaMemcpyAsync is supported in the new 2.0 SDK. However, I’ve heard from some colleagues that this will not actually run on current hardware.

2

In works on all compute 1.1 hardware, which is all CUDA capable hardware that is not G80 (8800 GTX, Tesla).

3

I have Async memcpys working on an 8800GTX, Compute Cap 1.0.

Streams do not work on the 8800 GTX, but async kernel calls and async memcpys do.

4

Yes, you can do async memcpys. But you can’t overlap them with kernel executions on 1.0 hardware which I assumed is what the OP was asking about.

5

I use async streams with 2x GF8500GT… but it’s curious… the cudaCreateStream() always returns a stream handle of “1”… even with multiple threads and GPUs plugged.

6

Thanks for all the help so far.

So let me try to see if I understand:

Compute 1.0 capability: Kernel will not execute while async memcopy is working

Compute 1.1 capability: Kernel WILL execute while async memcopy is working

Do I have this right?