PCIe x16 vs x8 vs x4
8 Feb 2011 23:23:20 UTC
Topic 195640
(moderation:
i had perform some tests and found, what if i plug GPU in PCIe x8 slot i get about 15% performance loss.
in x4 slot i get almost 40-45% loss.
is it ok or is it remnant of nvidia 260.xx bug workaround?
(gtx 560, phenom II 3GHz, linux x86_64)
thank you.
PCIe x16 vs x8 vs x4
)
I saw some performance loss going from an x16 slot to an x8 slot as well. I don't have the percentages but the time to complete each work unit went up some. Upgrading the drivers did not seem to make any difference. Based on that, I keep my cards in the x16 slots.
i have two, that's why i'm
)
i have two, that's why i'm asking. one in x16 slot and another in x8.
RE: i have two, that's why
)
In that configuration it is very likely (depending on your specific motherboard) that both cards are running as x8. With half the PCI-E lanes with which to move data back and forth between the GPU and CPU, it's not surprising that there is some loss of performance (per task) vs. a single card running as x16.
-- Tony D.
no. one is in x16 slot and
)
no. one is in x16 slot and another is in x8. slots on MB are marked and nvidia control center confirm it:
GPU 0: x16 lane
GPU 1: x8 lane
RE: no. one is in x16 slot
)
There probably is not much you can do to improve performance of GPU 1 due to board limitations. If you went with another board / chipset like the x58, which has 32-lanes available for graphics, you could give 16-lanes to each card.
Hmmm that's probably the
)
Hmmm that's probably the reason why my GPU took 38 minutes per CUDA WU.
But two take 58 minutes per WU !
Thanks for the information.
dunx
it is not PCIe speed related
)
it is not PCIe speed related issue. anyone running two WU gets such performance gain does't matter PCIe speed.
X58 for me ! Dammit I
)
X58 for me !
Dammit I will get these two cards running multiple WU's faster LOL
dunx
Sorry I wasn't clearly
)
Sorry I wasn't clearly stating that I was running two WU on TWO cards, so the speed difference is due to PCI-E sharing 16 lanes between two cards.
dunx
no, "Hmmm that's probably the
)
no,
"Hmmm that's probably the reason why my GPU took 38 minutes per CUDA WU.@
But two take 58 minutes per WU !"
this typicall result for 2 WU per one GPU.
for 2 WU per 2GPU you will get similar result as 1 wu per 1 gpu