There's a 60 MHz FSB option. Maybe the PCI bus runs at 30 MHz in this case, I'm not sure. There's no specific PCI bus speed option via jumper or BIOS. I think I should try the games in software mode and see if there is any change.

Maybe the PCI bus runs at 30 MHz in this case

No.

EDIT:
Ok, I did a fresh Win98SE install on my VS440FX board and will do some benchmarks with Quadro4 100 NVS too.

EDIT2:
Penium Pro 200Mhz 256Kb cache
80Mb RAM
Quadro4 100 NVS with Detonator 28.32
DirectX 7.0a

Quake 2 v3.20 without Vsync (640x480 16-bit, no sound, demo1.dm2) - 49.5 fps
Quake 2 v3.20 with Vsync (640x480 16-bit, no sound, demo1.dm2) - 39.8 fps
Incoming (1024x768* 16-bit, no sound, Vsync option does not change fps) - 40fps

So there's something strange with your Nvidia results and maybe other cards too.
*For some reason Incoming has slightly better results with higher resolution.

EDIT3:
MDK2 Demo without Vsync (640x480 16-bit, Hardware T&L, no sound) - 29.5fps
MDK2 Demo with Vsync (640x480 16-bit, Hardware T&L, no sound) - 29.3fps
MDK2 Demo without Vsync (640x480 16-bit, Software T&L, no sound) - 27.1fps
MDK2 Demo with Vsync (640x480 16-bit, Software T&L, no sound) - 26.97fps
Expendable (640x480 16-bit, no sound, Vsync option does not change fps) - 20fps

EDIT4:
I've noticed that all benchmarks were done in 1024x768 resolution, so I'll retest everything later to see if my results will be significantly lower. Although I doubt it.
Maybe will dig up my ESS1688 for consistent comparison with sounds enabled.

P.S.
BTW how to enable DMA mode for IDE on that motherboard? DMA flag always resets afrer reboot.
P.P.S.
Ok, apparently VS440FX does not like my IDE-to-SATA adapter or SSD itself. Bummer.

I didn't see your post for some reason. I'm not sure about the IDE DMA. I've had some boards that resets this check box upon reboot. I haven't dug into the reason or a fix.

Also, I found a small type in one of the data sets and added this footnote on one of the posts,

EDIT (25 Nov. 2019): I found a typo. For the game Shogo, the Savage4 framerate should be 27.6, not 25.15. Thus the Average for the Savage4 in the Combined Chart should be 29.9, not 29.7. In the Direct3D chart, the Savage4 value should be 28.9, not 28.7. The difference is small, so I don't plan on updating the charts.

I'm using the PPRO-512K at 233 MHz, so for consistent results, if you have this processor and can run it at 233 MHz, that could be helpful.

I have run all the games with the same resolution and settings as the PPRO using the PIIOD333 and am preparing a chart now.

Attached are the results of testing the same set of games using the PIIOD-333 with the Savage4. The data table is provided here:

I separated the 3D accelerated tests, that is, tests which utilise the graphic card's processing unit, from those of strictly software mode tests (CPU does all the crunching). This data has been charted and is provided here:

I then decided to look at the percent increase in performance per game for, both, 3D accelerated and software modes. These charts are provided here:

5 of the 15 accelerated games showed no distinguishable performance benefit when using the PIIOD over an ordinary PPRO at 233 MHz. Even the top teir performers of Half-Life, Expendable, Turok II, and Turok only showed a 15-25% improvement, with an average improvement of only 8.9% amongst all accelerated mode games tested here.

The benefit of using the PIIOD333 in software mode is more clear, with an average boost of 27.7% and a maximum boost of 33.41% in Shogo. This still seems smaller than what I'd expect from a PIIOD processor which runs 43% faster (100 MHz more) and has MMX, amongst other improvements.

I can understand the advantage of upgrading one's CPU to the PIIOD if their primary concern was for data crunching in software mode, but not so much for using 3D graphic card accelerated modes. The gains on said hardware would have been too slow to warrant the price to upgrade.

Makes sense really, since you're still limited to 66MHz FSB to DRAM and 33MHz PCI to VRAM

Penium Pro 233Mhz 256Kb cache
80Mb RAM EDO
Intel VS440FX
Quadro4 100 NVS with Detonator 28.32
Windows 98 SE (no additional updates)
DirectX 7.0a
===========================================================================
Windows:

Quake 2 v3.20 (without Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 53 fps
Quake 2 v3.20 (with Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 41.0 fps*
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, Medium geometry, no sound, four.dm_68) - 26 fps
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 24 fps
Incoming (1024x768 16-bit, no sound, Vsync option does not change fps) - 45.5 fps
MDK2 Demo (without Vsync, 1024x768 16-bit, Hardware T&L, no sound) - 32.27 fps
MDK2 Demo (without Vsync, 1024x768 16-bit, Software T&L, no sound) - 30.35 fps
Expendable (1024x768 16-bit, no sound, Vsync option does not change fps) - 22 fps

DOS:
Quake v1.06 (320x200, no sound) - 53.2 fps
Quake v1.06 (640x480 FastVid, no sound) - 22 fps

Judging from my results, I can assume that underclocked GeForce 2 MX without an optimal driver is the real problem in your tests.

According to Everest/Aida the chipset timings are optimal:
DRAM Read Burst Timing 2-2-2
DRAM Write Burst Timing 2-2-2
RAS To CAS Delay (tRCD) 1T

EDIT: Sidenote
Here's my Quake results on a PII system.

Pentium II 233Mhz Klamath
256Mb RAM SDR 2-2-2-6
ASUS P3V4X (VIA Memory interleave Enabler 0.15 patch, RivaTuner AGP4x fix)
GeForce 3 Ti 200 with Detonator 28.32
Windows 98 SE (no additional updates)
DirectX 8.0
===========================================================================
Quake v 1.06 (320x200, no sound) - 58.5 fps
Quake v 1.06 (640x480 FastVid, no sound) - 23.8 fps
Quake 2 v3.20 (without Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 71.5 fps
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 34.5 fps

P.S.
As mentioned before, screen resolution does not affect my results at all or the margin of error is 0.5fps.
*Major note: apparently Nvidia Vsync in Quake 2 is broken. Performance reduced, but screen tearing is still present.

Penium Pro 233Mhz 512Kb cache

===========================================================================
Windows:

Quake 2 v3.20 (without Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 56 fps
Quake 2 v3.20 (with Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 42 fps
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, Medium geometry, no sound, four.dm_68) - 28.3 fps
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 25.8 fps
Incoming (1024x768 16-bit, no sound, Vsync option does not change fps) - 49 fps
MDK2 Demo (without Vsync, 1024x768 16-bit, Hardware T&L, no sound) - 34.55 fps
MDK2 Demo (without Vsyncm, 1024x768 16-bit, Software T&L, no sound) - 31.92 fps
Expendable (1024x768 16-bit, no sound, Vsync option does not change fps) - 23.6 fps

DOS:
Quake v1.06 (320x200, no sound) - 57.2 fps
Quake v1.06 (640x480 FastVid, no sound) - 22.7 fps

EDIT: Also I've decided to directly compare my Pentium Pro results with Penium II/ASUS P3V4X combo, which has VT82C596B south bridge.

Pentium II 233Mhz Klamath
256Mb RAM SDR 2-2-2-6
ASUS P3V4X (VIA Memory interleave Enabler 0.15 patch)
Quadro4 100 NVS with Detonator 28.32
DirectX 7.0
===========================================================================
Windows:

Quake 2 v3.20 (without Vsync, 1024x768 16-bit, no sound, demo1.dm2) - 52 fps
Quake 3 v1.32 (without Vsync, 1024x768 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 25.8 fps
Incoming (1024x768 16-bit, no sound, Vsync option does not change fps) - 46 fps
MDK2 Demo (without Vsync, 1024x768 16-bit, Hardware T&L, no sound) - 33.81 fps
MDK2 Demo (without Vsync, 1024x768 16-bit, Software T&L, no sound) - 31.21 fps
Expendable (1024x768 16-bit, no sound, Vsync option does not change fps) - 24 fps

DOS:

Quake v1.06 (320x200, no sound) - 59.1 fps
Quake v1.06 (640x480 FastVid, no sound) - 23.5 fps

I can understand the advantage of upgrading one's CPU to the PIIOD if their primary concern was for data crunching in software mode, but not so much for using 3D graphic card accelerated modes. The gains on said hardware would have been too slow to warrant the price to upgrade.

I think S3 Savage 4 is slowing you down here. Because you're testing everything in 1024x768 resolution. And even if it's 16-bit color, S3 Savage 4 has only 64-bit memory bus. Pentium II Overdrive should scale better. For example, my Expendable result scale up to 21% with Pentium II 300Mhz on the same 66Mhz bus (29.1 fps), so yours should scale around 25-30%. In other words, each 33Mhz above PPro 233Mhz results should give 10% fps increase across the board.

Do you have a GF2MX to test?

The Serpent Rider wrote:

I think S3 Savage 4 is slowing you down here. Because you're testing everything in 1024x768 resolution. And even if it's 16-bit color, S3 Savage 4 has only 64-bit memory bus. Pentium II Overdrive should scale better. For example, my Expendable result scale up to 21% with Pentium II 300Mhz on the same 66Mhz bus (29.1 fps), so yours should scale around 25-30%. In other words, each 33Mhz above PPro 233Mhz results should give 10% fps increase across the board.

Except for Turok, which is at 800x600.

The Serpent Rider wrote:

Judging from my results, I can assume that underclocked GeForce 2 MX without an optimal driver is the real problem in your tests.

What are your results when using v45.23?

Do you have a GF2MX to test?

No. I probably can try to use AGP-to-PCI adapter and pray that it will work with some random GF2MX.

What are your results when using v45.23?

Didn't tried them on PPro.

Do you have a PIIOD333 to compare your Quadro results with on the same system?

Nope. Too rare and expensive for me. I probably can do more tests with Riva TNT2 PCI, but in 1024x768 it will be very noticeably slower.

Alright, well... I'm not sure what to conclude. Your results with the PPRO and NVS 100 are, on average, quite a bit better than all the cards I tested on the PPRO. I personally believe this is due to the newer and more advanced hardware on the Quadro rather than the driver release: 28.32 vs. 45.23. I had run through all the NVIDIA driver revisions when testing this GF2MX on my 486 and had settled on v45. Of course, I could test a few games using 28.32 on the PPRO if you think it will matter. And if so, which games in particular?

As I've said before, your GF2MX card has suspiciously low clocks. I would try to overclock it to 200/200 or more first. If that doesn't help, switch to Detonator 28.32 or 12.10. I think Quake 2, Turok and Incoming are good candidates for retesting,

Perhaps I am not clear on the objective. Are you anticipating that a GF2 will scale up better compared to the Savage4 when using the PIIOD?

Yes. PIIOD with overclocked GF2MX should probably scale more or less identical to PII/Quadro4 PCI. Full speed L2 cache can almost completely negate any advantages of more modern chipset.

I cleaned up my graphic card drawer and discovered that I have another GF2MX400. It is from PNY. I have a label on the back with my own handwriting:

1GeForce2 MX400
2NV11 
3B2
42.6 GB/s
5128-bit
664MB
7GPU: 200 MHz
8RAM: 167 MHz

It is the PNY card that I used in my 486. I don't know where this down-clocked piece of junk Jaton card came from that I used for the PPRO tests. I will need to retest the PPRO setup. Unfortunately, I now have socket 5 systems cluttering up my testbed for another build, Motherboard for socket 5 build , so I won't get to this for awhile.

The PCB datecode for the PNY is 45-2001, while the Jaton is 24-2003. Curious why Jaton bothered to produce such a slow GF2 - perhaps they were failed NV chips and were practically free. It will be interesting to see if the Jaton can clock to normal GF2 speeds, even with a fan installed. The Jaton comes fanless, but has a fan connector, while the PNY comes standard with a fan.

One more thing. Quadro 4 NVS 100 is apparently GeForce 4 MX 420, but with 64-bit DDR memory (128-bit SDR on 420). So it's only 1.4x-1.5x times faster than normal GeForce 2 MX 400, when CPU isn't a bottleneck.

And so I've decided to investigate further and to compare Riva TNT2 PCI on two different driver versions: 28.32 and 5.32. The goal is to see how much driver overhead much older version will allow to reduce.
Riva TNT2 PCI is a very weak card, with core and memory clocks rated at 125/150. So all tests were done in 640x480 16-bit resoluton to alleviate any GPU bottlenecks. Only exception - Incoming.

Penium Pro 233Mhz 256Kb cache
80Mb RAM EDO
Intel VS440FX
ELSA Riva TNT2 PCI
Windows 98 SE (no additional updates)
DirectX 7.0a
===========================================================================
Detonator 28.32:

Quake 2 v3.20 (without Vsync, 640x480 16-bit, no sound, demo1.dm2) - 38.8 fps
Quake 3 v1.32 (without Vsync, 640x480 16-bit, Lightmaps, Medium geometry, no sound, four.dm_68) - 24.0 fps
Quake 3 v1.32 (without Vsync, 640x480 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 22.1 fps
MDK2 Demo (without Vsync, 640x480 16-bit, Software T&L, no sound) - 23.82 fps
Incoming (without Vsync, 1024x768 16-bit, no sound) - 48.6 fps
Expendable (without Vsync, 640x480 16-bit, no sound) - 22.2 fps

Detonator 5.32:

Quake 2 v3.20 (without Vsync, 640x480 16-bit, no sound, demo1.dm2) - 43.2 fps
Quake 3 v1.32 (without Vsync, 640x480 16-bit, Lightmaps, Medium geometry, no sound, four.dm_68) - 25.3 fps
Quake 3 v1.32 (without Vsync, 640x480 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 23.4 fps
MDK2 Demo (without Vsync, 640x480 16-bit, Software T&L, no sound) - 24.81 fps
Incoming (without Vsync, 1024x768 16-bit, no sound) - 50.4 fps
Expendable (without Vsync, 640x480 16-bit, no sound) - 23.19 fps

Conclusion:
Early driver is moderately better in some situations and overall has better performance across the benchmarks. So if you don't have masochistic tendencies to play 3D accelerated games, released after year 2000 - pre 1x.xx Detonator should be your GO TO choice for all TNT-GF2 spectrum of cards. Especially if you're using CPU slower than a Pentium Pro or an early Pentium II.
That said, something like 3dfx Voodoo 3 PCI will still be better than any other contender, due to a superior driver/bus agnostic chip.

Penium Pro 233Mhz 256Kb cache
80Mb RAM EDO
Intel VS440FX
3dfx Voodoo 3 3000 PCI with official driver 1.07.00 WHQL
Windows 98 SE (no additional updates)
DirectX 7.0a
===========================================================================
Windows:

Quake 2 v3.20 OpenGL ICD (without Vsync, 640x480 16-bit, no sound, demo1.dm2) - 43.3 fps
Quake 2 v3.20 Glide (without Vsync, 640x480 16-bit, no sound, demo1.dm2) - 57 fps
Quake 3 v1.32 OpenGL ICD (without Vsync, 640x480 16-bit, Lightmaps, Medium geometry, no sound, four.dm_68) - 28.6 fps
Quake 3 v1.32 OpenGL ICD (without Vsync, 640x480 16-bit, Lightmaps, High geometry, no sound, four.dm_68) - 27 fps
MDK2 Demo OpenGL ICD (without Vsync, 640x480 16-bit, Software T&L, no sound) - 25.26 fps
Incoming (without Vsync, 1024x768 16-bit, no sound) - 51.2 fps
Expendable (without Vsync, 640x480 16-bit, no sound) - 25.89 fps

DOS:

Quake v1.06 (320x200, no sound) - 53.4 fps
Quake v1.06 (640x480 FastVid, no sound) - 22.3 fps

So looks like Quake 2 just loves raw GPU power, which is only partially limited by slow CPU and/or PCI bus. 3dfx MiniGL is an exception here, but image quality is worse.

So on average, you noticed about a 4.6% gain in gaming performance when using Detonator v5.32 over v28.32 on a TNT2 card with a PPRO-233 at 640x480. I'm curious if the percent gain is reduced at resolutions of 1024x768 or greater.