Testing PCIe At x16/x8 At Three Generations, From 15.75 To 2GB/s
Myth: PCIe x16 (Gen 3) is essential for realizing a graphics card’s top performance.
The ASRock Z87 Extreme6's UEFI includes a feature that allows per-slot configuration of PCIe transfer rates. You’d normally never want to change this, but for our tests, the feature is incredibly valuable.
The Extreme6, unlike the higher-end Extreme9, does not include a PLX PEX 8747, and is consequently limited to 16 lanes of connectivity. Furthermore, hardware strapping sets link width based on whether one, two or three graphics cards are present, automatically configuring the controller’s 16 lanes in x16-x0-x0, x8-x8-x0 or x8-x4-x4 mode.
We were thus able to test all three generations at x16 and x8 widths using enthusiast-class cards from both AMD and Nvidia. Just bear in mind that even when we're testing cards at x8, we're not testing CrossFire or SLI; we’re looking at single-card setups here.
Thomas conducted a similar test in 2011, the age of PCIe 2.0. Do his findings still hold true?
We tested using Unigine's Valley 1.0 benchmark. In each run, we let the cards reach their thermal throttling range and began testing once they were stably within a tight temperature/frequency range. We’re looking at real-world behavior; these cards are capable of higher frame rates, though they can’t be sustained over time under normal conditions due to thermal throttling.
As you can see, the GeForce GTX 690 achieves a 2.9% performance increase going from an eight-lane first-gen link (2 GB/s) to x16 PCIe 3.0 (15.75 GB/s), while the Radeon R9 290X (press version) manages a slightly more meaningful 6.4% gain. In both cases, the cards are not substantially bottlenecked by PCIe in Unigine Valley 1.0 down to 2.0 GB/s. That’s why the gains realized by ramping up to 15.75 GB/s are incremental at best.
One caveat is that we're not testing extreme setups with three or four cards in SLI/CrossFire. It would certainly be interesting to see if these results carry over to those multi-GPU setups. We're willing to bet, based on the bandwidth being utilized, that there would still be little to no change above an eight-lane PCIe 2.0 link. If you have your own concrete data, feel free to post a comment below!
PCIe Conclusions And Additional Thoughts
Even enthusiast-class graphics cards are not particularly bandwidth-hungry from a PCIe bus standpoint. Whether you use a PCIe 2.0, PCIe 3.0, a x8 link, or a x16 connection, for a single card, it essentially doesn’t matter to performance. If you have a third-gen x16 slot available, by all means use it. Just don't expect visible gains. And, if you’re using one card, don't consider a third-gen-compatible motherboard or processor a necessary upgrade based on PCI Express alone.
The story gets more complicated when you start taking multiple cards into consideration, and that’s where most enthusiasts probably want to see this experiment head. While the GeForce GTX 690 is, essentially, a pair of underclocked GK104s in SLI on a PCB, we can understand the desire for more exotic configurations. The link above from 2011 has some older data on SLI/CrossFire, which is worth looking at.