PCIe Bandwidth & SLi... I don't understand it, help?

DanJames93 · 2 Dec 2012 at 22:46

I'm in the process of upgrading my system. I am going from an AM3 Athlon II to a 1055 i7, and a GTS450 to a GTX570/80 or 670/80, possibly in SLi but I need some advice on PCIe bandwidths as I don't really understand them..

I'm looking at motherboards for around £120 that support SLi, but here's what's confusing me...

"Supports 1x PCIe 3.0 at 16x or 2x PCIe 3.0 at 8x"...

Won't one card running at 16x be the same as two running at 8x, as the motherboard is allowing half the bandwidth.

Or... would two at 8x be faster, because something tells me that PCIe 3.0 can support a much, much higher bandwidth than most (maybe all) modern GPU's can possibly support?

(I'm considering the ASrock Z77 Extreme4 by the way)... If anyone has suggestions, please keep it reasonable. I don't have stupid amounts of money to spend.

Thank you!

bremen1874 · 2 Dec 2012 at 22:51

DanJames93 said:
Or... would two at 8x be faster, because something tells me that PCIe 3.0 can support a much, much higher bandwidth than most (maybe all) modern GPU's can possibly support?

Basically this.

mickyflinn · 2 Dec 2012 at 22:53

Take a look at this.

http://forums.overclockers.co.uk/showthread.php?p=21534555

Martini1991 · 2 Dec 2012 at 23:00

Imagine it as a speed limit.
16X PCI-E 2.0 won't get saturated by any GPU now.
PCI-E 3.0 basically has twice as much bandwidth, but it's like saying "We can now go to 120MPH, but our cars only go 30MPH", and then the PCI-E 2.0 16X would be 60 MPH, PCI-E 2.0 8X would be 30 MPH etc.

DanJames93 · 2 Dec 2012 at 23:16

Martini1991 said:
Imagine it as a speed limit.
16X PCI-E 2.0 won't get saturated by any GPU now.
PCI-E 3.0 basically has twice as much bandwidth, but it's like saying "We can now go to 120MPH, but our cars only go 30MPH", and then the PCI-E 2.0 16X would be 60 MPH, PCI-E 2.0 8X would be 30 MPH etc.

Thanks for the replies..

So, lets say a GPU will only use 25% of the supported bandwidth of a PCIe 3.0 slot running at 16x, it would theoretically take a total of four GPUs running at 4x before the full bandwidth of the GPUs (and motherboard's PCI slots) are used to the full extent.

I get it now.

Martini1991 · 2 Dec 2012 at 23:21

Well, each slot can have their own set of bandwidth, so the 4 GPU's could all be using 25% of the available bandwidth available on their slot, I'm not sure such a board exists however

Natively, Ivy Bridge will support 2 PCI-E 3.0 lanes at 8x each (So each lane has PCI-E 3.0 8x bandwidth) You can then get chips that add bandwidth to the other lanes.

OliverDavis · 3 Dec 2012 at 09:04

Some slots share bandwidth. You have 2 PCIe 3.0 slots running at x8. If required you can have one running at x16, and the other deactivated.

Even a GTX690 Doesnt use all the bandwidth of a PCIe 2.0 x16 Lane. PCIe 3.0 has twice the bandwidth. So a PCIe 3.0 x16 Lane is equal to a PCIe 2.0 x32 Lane (These do not exist)

By running 2 PCIe 3.0 x8's You would not bottleneck (run out of bandwidth) with SLI 690's. A GTX 690 is equal to 2 GTX 680's.

So even with 2 GTX 680's in SLI at x8, each would only really need x4 to run.

So you would only be using 50% of your bandwidth or less. Its a shame you then wouldn't have any more PCIe Slots left (apart from x1) for other expansion cards.