I've seen the benchmarks, but I've also seen a discrepancy between the benchmarks of sites that I regularly use. Normally I wouldn't question the performance between AGP and PCI versions of the same part, but the VSA-100 chips of the Voodoo5 do not support AGP texturing, one of the main speed advantages of the AGP bus.
Anandtech shows the cards performing nearly identically, with the PCI slower by a couple frames, at most, in only instance.
Firingsquad, on the other hand, shows the PCI lagging behind the AGP by over fifteen percent in many instances.
They both used 98SE and the same drivers, final official drivers v1.01.00. Anandtech used an Athlon 750, 128MB PC133 SDRAM, and a 20GB 7200 RPM HDD; Firingsquad used a PIII 750E, 128MB PC100 SDRAM, and a 20GB 7200 RPM HDD. These are fairly comparable systems. Both included the same demo of Q3: Arena in their benchmarks using the same resolutions (although I'm unsure what quality the Anandtech benchmarks used).
I can only think of one thing that might explain the discrepancy between benchmarking efforts. Anandtech points out that if a previous 3dfx card in your system was not uninstalled properly (with the previous 3D adapter left behind, incorrectly identifying your new Voodoo5 5500 PCI as "3dfx Voodoo Series"), your performance can be significantly reduced. Perhaps Firingsquad did not uninstall their previous 3dfx cards properly.
Would anyone be interested in benchmarking both the PCI and AGP versions of the Voodoo5 5500 on the same hardware?