Pinned memory memcpy speed with 2 cards? pinned memcpy bandwidth drops to 50%!!!

Accelerated Computing CUDA CUDA Programming and Performance
1

Greetings,

Can anyone running 2 GPUs or a GPU and another graphics card for displaying post their SDK bandwidth results for pinned and pageable memory copy?

I have observed that if I use a single 8800GTS I get a significant speedup from using pinned host memory, as expected. However, if I enable a secondary graphics card, in this case a FX550, I stop getting any significant benefit from pinning. Has anyone observed any similar effect?

I am using the 6.14.11.6375 drivers on both cards.

Thanks,

Carlos

2

Isn’t that a motherboard problem? Does it support 2x16 PCI-E lanes or does it share when two cards are installed?

3

The motherboard has 2 PCI-E 16x slots but I can’t tell if it shares the bandwidth or not. Is there a way to check this programatically other than looking up the motherboard specs? Is there a whay to get this information from the “system information” in Windows XP or through some system call?

Thanks!

Carlos

4

I have used a G80 / G86 combination or a G80 solo. I don’t see any benefit from pinned memory with the ‘second’ card (the G86) but it has had no impact on the first card, which (still) benefits substantially from pinned memory.

Geoff.