Using different chips when comparing the buses is apples and oranges. The Riva128 was an excellent DOS chip, where your newer chips probably weren't designed with DOS performance in mind. If you compared it to an AGP Riva 128 you could draw conclusions about the impact of the bus, right now you just have two unknown factors.

I also doubt the extra bandwidth or other features of AGP would be utilized under DOS as they require separate drivers to work. In any event the assumption that more bus bandwidth = more performance is flawed. *If* the bus is the primary bottleneck that would be partially true, but most of the time the processing by the GPU or the bandwidth between video memory and GPU is far more relevant. It does depend on the workload, but usually the AGP version of an early AGP chip was only a few % faster than the PCI version in most benchmarks or games, and that was usually more due to lower latencies than higher bandwidth as such.

Run FASTVID or MTRRLFBE to enable fast VGA video and then redo the benchmarks.

You can get these in the Phils pack:
https://www.philscomputerlab.com/dos-benchmark-pack.html

Have you tried using fastvid or some other video optimiser on the AGP system.

Often, the AGP bus is not configured properly for maximum DOS performance by the motherboard on bootup.

One of the AGP-cards that is very cheap and very fast under DOS is the Matrox G100.

It beat all other cards I tested(Riva128 PCI/AGP, Ark2000 PCI, Matrox Millenium II PCI, Riva TNT PCI, GeForce2MX PCI, TNT2 PRO AGP).
Albeit I do not have too many cards so take this verdict a huge grain of salt. I recently got a Matrox G400 AGP and it is even slightly faster than the G100 AGP in 2D/buffer operations.

But, as already said, use fastvid if you haven't already and retest.

The AGP cards you tested are from exactly the era where simple, brute force VGA/VESA speed was not as important anymore... who bought such a card to run Quake, Blood or another classic in high res only to score peak frame rates?

ATi Rage 2 PCI and AGP seem really fast as well ! Speedsys reports over 70MB/s of bandwidth with PCI bus or AGP (with FASTVID)

Thanks for the replies. As stated in my original post, I have MTRRLFBE in my autoexec.bat (MTRRLFBE VGA WC and MTRRLFBE LFB WC). Without that, I'd get maybe half the FPS, if that.

I dug up my old Asus P2B and will retest using that. I have a hunch that the VIA Apollo Pro 133 chipset may be a bit crappy. I also have some AGP G100 (and maybe G200?) cards in storage that I can try. Forgot to ask, what kind of FPS should I expect from Quake 1 in 640x480 on a P3-800?

When I tested AGP 2x/4x/8x boards for raw bandwidth under DOS, I got essentially the same results on all of them. So I assume the higher speeds aren't even enabled when booting to DOS. None were slower than PCI though!

AGP is designed to always use burst transfers, unlike PCI. Maybe Quake is doing something weird like writing pixels out of order or reading them back for some reason.

Quake is more limited by the cpu speed than the graphics card - it doesn't seem to do any magic to the card except uploading the final frame. Doom seems different in this regard.
But, no guarantee.

I still remember '96 and the disappointment on peoples faces when their beloved Matrox was only a hair faster than a Trident card in Quake, all because the frame rate was so low that no high bandwidth was required to show those sappy 10-15fps in 320x200.
There was no concept in people's minds that the card in itself couldn't accelerate anything because it was just used as a frame buffer... though one got the impression via ads that all that "accelerator magic" was, maybe would, support any game out there... somehow.

As frame rates go up the raw transfer rate to the frame buffer gets more important. A typical PCI VGA is no bottleneck for a 486 or a slower socket 5/7 system but makes a huge difference for that p3-800. Especially write combining, or the lack thereof which can really cut frame rates in half.

When you actually play it won't harm 320x200 but in 640x480 and above this effect shows up at normal frame rates, not just in timedemo. And there are other strange factors too, like GUS Classic sound streaming magically stealing 30% of performance on many chipsets.