That's more bandwidth than some GPU's. DRAMless GPU's incoming? (Unlikely, bottlenecked by System Memory.)
Yeah, that's been tried. Back in the late 1990's, there was an idea that AGP was fast enough to let GPUs use system memory. I think the GPU still had DRAM for the frame buffer, but not much more than that. It didn't work terribly well, IIRC.
What would kill you is latency. DRAM latency on GPUs is like 200 - 300 ns, but for them to access memory connected to the host CPU would probably almost double that, and could be even worse under high load.
Another bad thing about hitting PCIe that hard is that it'd use more power than accessing memory local to the GPU.
Lastly, it'd be competing with the CPU for memory bandwidth, and thus bogging down any parts of a game engine that are CPU-limited. We already see this with iGPUs, but it'd be even worse with a dGPU.
For the average consumer it's pretty useless at this stage outside of low-PCI-E lane devices like SSD's.
Yeah, the SMI CEO just said:
Just to be clear, he's talking about PCIe 6.0 and you're already thinking about PCIe 7.0. I do appreciate your little thought experiment. But, by the time it's even a relevant question, the bandwidth needs of GPUs will probably have increased to the point where it would make no more sense than it does with PCIe 5.0, today.