4

I expected that GTX 680 (which is one of the latest version of GPUs) is capable of concurrent data transfer (concurrent data transfer in both direction). But when I run cuda SDK "Device Query", the test result of the term "Concurrent copy and execution" is "Yes with 1 copy engine", and it means that the GPU can not do concurrent data transfer.

I wonder if my testing result also happens to you? And can you share with me which devices are capable of concurrent data transfer?

Thanks!

einpoklum
  • 118,144
  • 57
  • 340
  • 684
Blue_Black
  • 307
  • 1
  • 3
  • 11
  • 1
    FWIW, "Yes with 1 copy engine" does not mean that "the GPU cannot do concurrent data transfer". Yes means Yes. It means it can do a data transfer at the same time as running kernels. If it said "with 2 copy engines" it would mean it can do bidirectional transfers while also running kernels. – harrism Aug 29 '12 at 03:37
  • @harrism, thanks for your comment. But when I say "concurrent data transfer", I do mean bidirectional transfer while also running kernels. I got this name from chapter 3.2.5.4 of CUDA_C_Programing_Guide V4.2 – Blue_Black Aug 29 '12 at 21:48

1 Answers1

4

Dual copy engines are available on Tesla cards and modules:

http://www.nvidia.com/object/why-choose-tesla.html

http://www.nvidia.com/docs/IO/43395/NV-DS-Tesla-C2075.pdf

Also, some Quadro models provide dual copy engines, e.g.:

http://www.nvidia.com/docs/IO/40049/Dual_copy_engines.pdf

http://www.nvidia.com/object/quadro-k5000.html

njuffa
  • 23,970
  • 4
  • 78
  • 130
  • Thanks for your answer, njuffa! But one more question is that will concurrent data transfer double the total bandwidth? In other words, will backward transfer limit forward transfer speed if they happens concurrently? – Blue_Black Aug 28 '12 at 19:52
  • PCIe is full-duplex (see http://en.wikipedia.org/wiki/PCIe), so simultaneous upload and download can both get the full bandwidth. However, you may run out of system memory bandwidth when using both copy engines, especially with PCIe gen3 (Kepler-class GPUs). – njuffa Aug 28 '12 at 21:54