If you will have to push data from system memory to VRAM on GPU you will experience sweet swapping and 1-2,5 fps average. GPU memory bandwidth is measurement of data transfer between GPU core and Video card RAM. System memory is too slow and once you have not enough "buffer" you are going to struggle with 20GB/s system memory bandwidth.
I think you're confused a bit, or maybe I am. Data is always being
swapped out between the hard drive, system memory and VRAM (in that order) with the CPU controlling everything.
Modern graphics engines stream or access data constantly in the background from the hard drive/SSD to system memory, and finally to VRAM so the GPU can render it.
As long as everything is balanced and you're not playing at settings that are too high for your system, there should never be a time where you run out of frame buffer.
Of course, there's been times where an engine or API can be horribly inefficient when it comes to memory usage, but game engines nowadays are much better at doing that now that there's DX11 and the techniques are better.
The effect is the same when you don't have enough system memory for a game, even though you have enough VRAM.. Instead of pulling data from the system memory which if
WAY faster than the hard drive/SSD and moving it to VRAM, the CPU has to access the hard drive directly which slows everything down to a crawl..
Those graphs prove what I've been saying more than anything. Modern games have very efficient memory access/usage algorithms that STREAM data constantly, which reduces the overall memory footprint. (I don't think those graphs measure system memory either, but VRAM)
Normally this level of detail would require many gigabytes of system memory and GPU VRAM, but thanks to a fairly efficient streaming system Last Light’s world uses less than 4GB of system memory, and here's a shocker, less than 2GB of VRAM, even at 2560x1440 with every setting enabled and maxed out. We had a hard time topping even 1 GB of VRAM usage