GTX 680 and GTX 780 with only one copy engine?

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m starting to experiment with CUDA streams and asynchronous data transfers, but I’m not getting the expected results on two systems I’m using, one with a GTX 680 (drivers v319.32, Ubuntu 13.10 64bit) and one with a GTX 780 (drivers v319.37, Ubuntu 10.04.4 64bit). In particular I’m not able to overlap host-to-device transfers with device-to-host transfers, in fact the asyncEngineCount property is equal to 1. Even the example from [url]https://devblogs.nvidia.com/parallelforall/how-overlap-data-transfers-cuda-cc/[/url] behaves as I’m using a C1060 with only one copy engine. Is this possible? Do these two cards have only one copy engine or the drivers are too old?

Thanks in advance

2

The GTX 680 and 780 both have only one engine as you’ve found. Two DMA engines enable simultaneous asynchronous memory copies. That’s a feature of the professional Tesla GPUs, not the consumer ones.

3

Ok, so I guess there’s nothing I can do for this… thanks for the reply

4

Just checking you know that your cards can overlap kernels with transfers. This is what your common or garden variety CUDA-elf means when they talk about “using streams for asynchronous transfers”.

5

Yes, this is what I’m achieving right now:

786×69 25 KB

I guess it’s the best with just one copy engine