Yup 128bit is the one to have, clocks can always be pumped a bit, but you're stuck with the bus width.
Closest thing really to a GF 3 PCI is a FX5200 PCI... it's a much maligned card, but knowing it's strengths and particular weaknesses can find a notch where it's useful. Most stuff it's always faster than 8500 and relatives, sometimes not by much. On "ATI favoring" DX 9 things, that's it's worst, and where all the FX cards drag down, and it looks bad against an MX460 ... though that would only be doing DX7 in the same game. However, on "neutrally programmed" or nvidia favoring DX9 it can actually beat the lower members of the 9600 line. DX8 stuff it's usually okay on floating between 8500 and 4200TI, on one or two titles it seems to pull on the 4200 a little. DX7 performance was not that good with early drivers, though I think things improved later, gets around 5000 in 3DMark2000 which is what you get with a Voodoo 5. Anyway, hardly anything where you could say "If there was a PCI GF3 it would have been faster". It launched into a time where games got rapidly more brutal on GPUs, even the top DX9 cards of ATI then had a really short life as adequate for AAA games, general recommendations for "XP gaming rig" tend to hit 3 and 4 generation newer cards and divide these back into "high end win98". So buying one at the time might have left you feeling rather cheated as it was slowest in everything getting released, but the dudes who got top of the line weren't making 60fps in much after a couple of months either. Going back in gaming from there though, it's that or the R200 PCI if you're stuck with PCI and next notch down is an MX PCI or a Radeon 7000 with the T&L crippled. And of course AGP can fly higher for the 2000-2003 notch.