ConclusionLooking at the results, we can see a whole lot of nothing. PCI-Express 4.0 achieves only tiny improvements over PCI-Express 3.0—in the sub-1-percent range When averaged over all our benchmarks, we barely notice a 1% difference to PCIe Gen 3. I also included data for PCI-Express Gen 2, data which can be used interchangeably to represent PCIe 3.0 x8 (or PCIe 4.0 x4). Here, the differences are a little bit more pronounced, but with 2%, not much to write home about, either. These results align with what we found in previous PCI-Express scaling articles.
That's of course a good thing as it confirms that you do not need an expensive PCI-Express 4.0 motherboard to maximize the potential of AMD's new Radeon RX 5700 XT. It also produces strong evidence that PCIe 4.0 won't be needed for even more powerful next-gen graphics cards because our three tested resolutions reveal more details.
If you look closely, you'll notice that lower resolutions show bigger differences in performance when changing the PCI-Express bandwidth, which seems counter-intuitive at first. Doesn't the graphics card work harder at higher resolutions? While that may (mostly) be true, graphics card load does not increase PCI-Express bandwidth; it actually lowers it because frame rates are lower. The amount of data transferred over the PCIe bus is fairly constant—per frame. So if the graphics card can run at higher FPS rates because the resolution is lower, the PCIe bus does have more traffic moving across it.
So even if next-gen graphics cards significantly increase performance, we won't see huge differences in PCIe requirements because you'd not use those graphics cards at 1080p, but rather 4K. In such a scenario, with FPS increasing on 4K, the difference in scores would be more similar to this review's data for 1440p, or even 1080p.
When looking at individual game results, the effects of constrained PCIe bandwidth vary wildly. Some games, Shadow of the Tomb Raider, for example, barely show measurable differences, while titles like Rage 2 and Wolfenstein are much more dependent on PCI-Express bandwidth. I can't see a clear trend between APIs or engines as it rather looks like a dependency on how the game developer chooses to implement their game and how much data they copy from the CPU to the GPU, or even back.
These results are also good news for people who consider running their graphics card at reduced link width, like x8 or even x4, to free up precious PCI-Express lanes for other devices, like storage.