It could be a fault of the morherboard design that causes the instability. The 3,3V regulator of the board might have a problem with a power hungry card like a GeForce2 GTS. Even some ATX boards of that era had this issue as they also used their own 3,3V regulator instead of using the 3,3V line from the PSU. Forcing PCI mode might put less stress on the card and could be the reason that it works. It would be interesting to see if a lowee power card like the GeForce2 MX would run stable in AGP mode.

Matze79 had similar problems and found a fix:
ALI Chipset vs AGP Card - K6-2 Commate S7AX
He disabled Sideband Adressing (and repaired broken solder joints on his board). Given going into PCI mode disables Sideband Adressing as well you should try that.

Yeah, I read that thread yesterday, but for some reason it didn't work for me - I guess AGP cards like Geforces are just too much for this motherboard to handle. I had a spare MX 400 laying around, so I decided to test this one as well.

Unfortunately things didn't go too smooth. Here are the results of benchmarks I tried to run:

3DMark 2000:

Geforce 2 GTS:
Crashed on every single one, regardless of settings

Geforce 2 MX 400:
1262 (AGP x1, SBA off, GAT 0)
Crashed (AGP x1, SBA off, GAT 1)
Crashed (AGP x2, SBA off, GAT 0)

Voodoo 3 2000:
1371 (Turbo settings in AGP tool)

3DMark 99 MAX:
(amusingly all cards worked rather well on this one with SBA off and AGP x1)

Geforce 2 GTS:
3213 (AGP x1, SBA off, GAT 0)
6018 CPU Marks

Geforce 2 MX 400:
3156 (AGP x1, SBA off, GAT 0)
5898 CPU Marks

Voodoo 3 2000:
2546 (turbo settings in AGP tool)
6044 CPU Marks

---
So yeah, here you have it. I'm still baffled why 3DM2000 crashed most of the time, while 3DM99 worked really well. Any further ideas?

Hm, maybe it has something to do with the inofficial K6 plus BIOS?

Kurasiu wrote on 2021-04-16, 09:46:

Then I found out forcing PCI mode actually makes the card stable... Which was to be expected, but completely pointless otherwise. It makes V3 run circles around the G2 (837 in 3DMark, compared to Voodoo's 1365)

After I had the artifact issue on Voodoo 3 (which apparently can be fixed, see here: Voodoo 3 2000 Artifacts and noisy video. ), I went back to testing other options and that also included my GF 2 MX400 and applying what I read in the meanwhile about this rather problematic board and first thing I tried was enable PCI mode via agp utility 140

After your thread I expected it to be a total disaster. But the opposite was the case. It was actually MUCH faster than the Voodoo 3.
GLquake didnt work obviously. But the rest is just crazy. Also the values for
Bus Transfer Rate & FAGP Test 16mb are crazy high now. I that because they are "broken" now? Or is the PCI slot just that much faster?
Also I realized that on the GL XS Benchmark I actually didnt have a big in the first loading screen (bar not showing), and the "scene 3" with the "Facescape" (high poly count test) also didnt have graphical glitches. Obviously stuff requiring glide wont work but I expected totally crushed performance. I guess it would be better on AGP, but still.
Then I also ran AGP on the settings you provided (which were the only ones which worked). My results are far better than yours though (and also better than the PCI ones). Basically outperforming the Voodoo 3 everywhere by 50-100%

As mentioned Glide wont work. And question is also whether these results actually translate into game performance. The Unreal Time demo is actually doing worse on direct3D than it does on the Voodoo 3 Glide (and OpenGL crashes) and its missing the reflections.

(These results for the GF have not been tweaked and I didnt disable vsync like I did with with the voodoo)