Well, that's Tom as we know it
Anyway, it largely depends on the Application you throw at it.
I mean, run Unreal Tournament on 2x/4xAGP vs. 8xAGP Cards, and you'll see those 1-2% (at best).
Do the same using the (extreme by all means) CodeCreatures benchmark, and you can see upto 60% in isolated Cases.
The trick is to force the Card into AGP Texturing, that's it.
But still, the usefulness of 8x AGP is (today) mainly limited to getting T&L data and other misc. stuff over the Bus to the Card quick.
AGP Texturing is really not used alot (if at all), since it drops performance way too low.
But if you want to see 8x AGP at it's best, use a 64MB Card, set it to 2xAGP and run UT2003 in maximum Texture Detail (requires manual Engine flush to enable true 256MB Textures); 128MB (default maximum) should so as well however.
Then, repeat procedure on 8xAGP.
(Problem here is finding a 8xAGP capable Card that actually still supports 2xAGP)
----------------
The difference from 4xAGP to 8xAGP in turn, is neglectable, simply because 4xAGP already almost exceeds most system's total available "spare" RAM bandwidth (which is naturally shared among other Tasks, e.g. PCI Bus & CPU).
So 8xAGP would just bottleneck the RAM even more, yielding either no reasonable performance increase (worst case), or still less than expected...
It's just hard to find the right combination of Cards and Applications that posess enough Textures to force AGP Texturing, and provide a Platform that can spare sufficient RAM bandwidth for the 8xAGP to actually use as advantage over 4xAGP.
-----------
But by all practical means, I agree :
We don't really need 8x AGP as of now.
And when the time comes we could actually need it, we'll find the performance drop being probably too significant to be useful.