PCI-e Lanes - a question.

[escribblings]

I am finally approaching getting a MB with PCI-e

Yes, you read that right - I am losing (or going to lose) my PCI-e virginity.

And I have a question.

I have seen some boards marked as say s PCI-e 16x/8x8x. Now I am presuming that this means there are 16 lanes shared between the 2 PCI-e sockets, and if only 1 is used it gets all 16, but if both are filled they get 8 each.

Have I got this right?

A sub question - which is more geared towards graphics cards - is does the number of lanes actually make a real world difference?

If I have my above assumptions correct, the P8Z77 WS appears to have 32 lanes shared between 4 PCI-e 3.0 sockets, allowing for configurations of 16/16/0/0, 16/8/8/0 or 8/8/8/8. Where as the Z77 Sabertooth appears to have 16 lanes shared between 2 PCI-e 3.0 sockets allowing for 16/0 or 8/8 configurations.

Now here is the question - does it really make any difference?

Lets assume that the 2 boards are practically equal performance wise in every aspect apart from PCI-e (I haven't looked into the benchmarks so this may not be true - but lets assume for the time being).

now lets also assume I have 2 identical GFX cards - I will take the GTX 690 as an example as it is one of the most powerful out there.

If I install both of them - would there be any difference between a board that can only offer 8/8 to a board than can offer 16/16?

If not, how far away are we from the GFX cards needing more than x8?

Thank you.

 

[rusabus]

Honestly, it doesn't make much difference. Especially with PCI-e 3.0. Toms has done several reviews with various configurations, which although a bit dated are still relevant:
http://www.tomshardware.com/reviews/pci-express-scaling-p67-chipset-gaming-performance,2887.html
http://www.tomshardware.com/reviews/pci-express-scaling-analysis,1572.html

The more recent of those articles uses PCI-e 2.0, which is half as fast as 3.0, so a 2.0 x16 slot is as fast as a 3.0 x8 slot. Even at that, an x8 only had a 2% hit on performance vs. x16.

In my opinion, if you're going for a multi-GPU setup, you'll do better going with a cheaper MB that offers x8/x8 connectivity and putting the extra $50 into the graphics cards. Naturally, x16/x16 will be a little better, but not as much better as $50 more spent on video cards would be. I guess what I am really saying is that I would not go out of my way to spend extra on x16/x16 - especially if the board/CPU support PCIe 3.0.

--Russel

 

[Pinhedd]

Have I got this right?

Yup! I assume that you're looking at a Sandybridge or IvyBridge processor. Each of these (with the exception of the high-end Sandybridge-E processors) has 16 PCIe lanes which are pulled out directly from the CPU. They can be configured as either a single 16x lane, or a pair of 8x lanes. The Ivybridge processors upgrade this to PCIe 3.0 and add an additional configuration of 8x/8x/4x for triple GPU configurations.

A sub question - which is more geared towards graphics cards - is does the number of lanes actually make a real world difference?

The number of lanes makes a small difference when it drops to 4x at PCIe 2.0 speeds or below but only on top end cards. There is no noticeable difference between 8x and 16x lanes running at PCIe 2.0 speeds, and 4x, 8x and 16x lanes running at PCIe 3.0 speeds.

 

[manofchalk]

Yep youv got that right.

As long as there is at least PCI-Gen2 8x bandwidth on a graphics card, it will not see any bottlenecking. More lanes and gens higher than that wont equal more performance except in some highly synthetic benchmarks.

Also if you plan to run cards in Crossfire/SLI, make sure every card has adequate bandwidth. The performance of the slowest card will determine the performance of the whole array, so hobbling a card with 4x bandwidth will likely lead to less performance than a single card.