Ok
@dullard Desktops w/ SLI / 2-3 GPU's in the case is a different story.
This first hurdle is the slots / speeds as most MOBO's hack off the speed of the slots as you use more than a single GPU.
X16 - slot 1 might be Gen3 or 4 with full x16 speed
X16 - slot 3 is chopped to x8 for bandwidth and reduces slot 1 to x8 due to the # of PCIE lanes available
If you happen to be using slot 5 there's a good chance it's running through the
DMI which in anything prior to ADL is 1/2 the speed DMI 3 (prior to ADL) and DMI 4 ADL+
The problem with DMI is it's sharing all of it bandwidth across multiple functions from drives / sound / GPU's / basically anything in the system that doesn't
require CPU direct lanes.
-------------------
Beyond this though with PCIE gen 5 / NVME around the corner it could consume your only PCI slot for a gen 5 adapter / NVME that can push 14GB/s. GPUs will take awhile to rev up their engines to get those sorts of BW to render even faster but, they're in the pipeline.
Do you need Gen 5 speeds on a NVME? No, most of the time even on a Gen3 drive it sits idle while the game is open / running since it's put into the RAM space after the initial load. The speed benefit comes with the the transfer from sotrage to RAM. Reducing the time spent waiting for the system to load things is where the improvements are.
Things are changing yet again though with MSFT testing Direct Storage that's supposed to streamline the process even further.
Taking a high level view to see the forest to see where you want to plant your trees helps get the best / most performance out of the HW you're putting into a system.