Right now AGP 4x is nearly pointless, the biggest performance boost comes not from the bandwith but from fast writes and other AGP 4x features. There are, however, a few applications that do utilize AGP 4x fully and right now will utilize AGP 8x fully: CAD/CAM/CAE and DCC. Using usually the AGP Pro Wildcat cards with textures that exceed the onboard memory of even the 384 Mb Wildcat 4, they have optimized drivers for hitting to main memory. They will notice an immediate difference from AGP 8x over AGP 4x.
As for DX9 using it, nah, still won’t really use it. DX9 is just DX8 but more programable to SAVE bandwith (just a little bit, of course). Ditto with OpenGL 2.0, which 3dlabs is the driving force behind. Heidi is right now possibly the only acceleration method that would benefit immediately from AGP 8x at all due to its high bandwith consumption (although it is a ton easier to program for). FYI: that’s Autodesk’s 3d acceleration API.