Intel should include a L4 cache on all cpus then!
Joking aside, the test you provided is very interesting, but still it may not be 100% correct.
They are simulating PCI express 2.0 speeds by a BIOS selection, while the signaling is still PCI Express 3.0.
PCI express 2.0 has a 8b/10b encoding scheme while PCI express 3.0 employs a 128b/130b encoding scheme. From that alone, the PCI-e 2.0 has an overhead of 20% while PCI-e 3.0 only 1.54%.
Data transfers have not only to do with bandwidth alone.They have to do with what data fragments we are dealing with too.
So in our 1080p testing, there could be a lot of IO with small fragments, that was handicapped due to the Sandy Bridge's PCI-e 2.0 controller and its 8b/10b encoding. The part we tested shows very little graphics anyhow, so IO could be hammered a lot more than the bandwidth itself.
It's like copying a single 100GB file from one place to another or copying the same data size in 100.000 files. It's not the same thing. Since the GTX 1070 has 8GB of video ram, which is more that enough to store all data required for the charlie-dont-surf-ometron and more, we may very well have to deal with many small data fragments, that hit that 8b/10b encoding wall.
At least that's my theory. I cannot be certain.
One other theory that could explain a cpu limit, is that maybe Modern Warfare Remastered employs complex AVX commands, that the Sandy Bridge must execute in two cycles and Skylake in one, so even at 3GHz vs 4.5Ghz it would still be faster. These AVX commands may have not been used in the cpu tests we did earlier and hence we did not get the whole picture.
Once more, having 1-2% difference in GPU load and 25% difference in performance, with the same card, same settings, same res, same location, is very weird. I cannot just accept it's a cpu limit (in processing power terms), without looking further.