Conclusion

Almost 2,000 individual benchmark runs later, we have a much clearer picture of PCI-Express scaling using the latest and greatest hardware. Our results here can also be extended to other platforms that use PCI-Express for graphics card connectivity, because performance of the PCI-Express controller itself can not lead to significant differences, as long as transfer rate and number of lanes remains the the same.

• The new PCI-Express 3.0 interface can provide around 1% performance boost for both HD 7970 and GTX 680. While this confirms that both cards provide working support for Gen 3, such a small improvement is clearly not worth worrying about. It certainly does not warrant buying a new processor or motherboard. PCI-Express is forward and backward compatible, so any PCI-Express graphics card will work in any motherboard's PCI-Express slot, no matter which version each component supports.
• PCI-Express 1.1 x8 poses a significant performance loss, so for people clinging on to old platforms, it's time to upgrade.
• PCI-Express configurations that promise the same performance, do deliver on it. For example, we see 1.1 x16, 2.0 x8 and 3.0 x4 with the same performance, within a 1% margin, which is not significant beyond benchmarking.
• Our testing confirms that modern graphics cards work just fine at slower bus speed, yet performance degrades the slower the bus speed is. Everything down to x16 1.1 and its equivalents (x8 2.0, x4 3.0) provides sufficient gaming performance even with the latest graphics hardware, losing only 5% average in worst-case. Only at even lower speeds we see drastic framerate losses, which would warrant action.
• Each game has different requirements for PCI-Express bandwidth, depending on the game engine design. Alan Wake is most dependent on a fast bus interface, losing up to 70% framerate, whereas Aliens vs. Predator handles bandwidth starvation the best, losing only 10% in worst case (1280x800 GTX 680).
• Contrary to intuition, the driving factor for PCI-Express bus width and speed for most games is framerate, not resolution. Our benchmarks conclusively show that with higher resolution, the performance difference between PCIe configurations shrinks. This is because the bus transfers a fairly constant amount of scene and texture data - for each frame. The final rendered image never moves across the bus, except in render engines that do post-processing on the CPU, for example Alan Wake. Even in that case, the reduction in FPS from higher resolution is bigger than the increase in pixel data.
• NVIDIA's GeForce GTX 680 suffers a relatively bigger performance hit from a slower PCI-Express interface than AMD's HD 7970. Going from x16 3.0 to x4 1.1 causes the HD 7970 to lose 14%, GTX 680 loses 27% real-life performance for the same transition. A reasonably accurate rule of thumb is that GTX 680 loses twice the percentage from slower PCI-E speeds, compared to HD 7970.
• PCI-Express 2.0 x8 is still a viable mode for 2-way multi-GPU. This is the mode most Core "Sandy Bridge" platform users will end up using for multi-GPU, and differences between PCI-Express 2.0 and 3.0 x8 is just 4% and 2% for the GTX 680 and HD 7970, respectively.
• PCI-Express 3.0 x4 is a revelation. Although we knew that on paper it provides bandwidth comparable to PCI-Express 2.0 x8, we were skeptical. The mode's real-world performance proves the theory, and could be a pleasant data point for users of performance and high-end Intel Z77 motherboards in the ATX form-factor, running Ivy Bridge Core processors, which have a third PCI-Express 3.0 x16 (electrical 3.0 x4) slot wired to the CPU.