AaronAsh wrote:

So those are immediately kind of interesting, as the Voodoo2s aren't really showing the kind of scaling you might expect and the TNT2 actually loses some performance going to a lower resolution! That to me suggests something in the system is getting swamped by the graphics cards output and totally failing to cope with it.

CPU is running at 5.5 x 100 - I did a run of SuperPi and it took 5:35; seems reasonable?

I did initially ignore the lower performance as me overestimating how well these machines used to run, but there are plenty of benchmarks of Voodoo2s/TNT2s with K6-2/3 CPUs which indicate this performance really isn't what you'd expect.

your superpi time is normal, so the problem is not within the cpu-cache-ram subsystem. now there are a few things to check out:
1. use everest or sisoft to see your tnt2 running in agp or pci mode.
2. your nvidia driver version? anything newer than 30.82 is not recommended on socket7 platforms.
3. check out the options in video driver: turn on optimizations, and disable vsync(may need the "coolbits" registry).
4. do you use built-in demo1 or crusher demo for quake2 benchmarking? i got 68fps with geforce2ultra in crusher, so if you are getting 40 crusher fps with tnt2 its probably right, but 40fps in demo1 is too low.
5. benchmark with sound disabled, running with sound can slow down by a lot. to disable sound, enter:
s_initsound 0
snd_restart
6. about that agp-aperture size, my suggestion is 64mb, setting it too small can reduce performance.

Didn't get much time to look at this over the weekend, but I did do some tweaks and also ran 3DMark2000

So I found a BIOS setting that has increased the AGP cards performance significantly - it was called "AGP Read Burst", and defaulted to disabled. When enabled it added around 10 fps to Quake 2s demo1 frame rate and around 300-400 to the 3DMark score!

Here are the updated benchmarks with it enabled:
TNT2 V770 AGP
(using the latest Diamond drivers which are based on the Nvidia 3.68 drivers)
1024x768x16
3D Mark 2000 - 1785
Quake 2 demo1 - 52

2 x Voodoo 2 SLI
(using the latest FastVoodoo drivers)
1024x768x16
3D Mark 2000 - 1216
Quake 2 demo1 - 33

So the AGP result is much better now, though still not in the 60 fps range in Quake 2 demo1 like I've seen other benchmarks achieve with similar CPUs.

The SLI Voodoo2 result is still strange - I actually think what's happening here is some problem with SLI itself; if I run Q2 demo1 at 800x600x16 I actually get a slightly higher average fps (36 vs 33) with SLI disabled! The system suggests SLI is working correctly - it shows up as enabled in control panel, and allows me to access 1024x768 resolution on the V2s, but I wonder; I did follow a guide to make my own SLI cable, so perhaps that is not working correctly. Still, according to page 8 of Phil's very handy Voodoo SLI scaling test (pdf here) I should still be seeing at least mid 50s frame rate in Q2 demo1 with even only a single Voodoo2 on a comparable system.

So I've made a bit of a breakthrough with this system. I picked up a really cheap Geforce 2 GTS on ebay, and decided to give that a shot in the system and discovered that, not only is it significantly faster than the TNT2 (of course) but it also nearly doubled the performance of the connected Voodoo2. I don't have the precise numbers to hand, but the Voodoo2 3DMark2000 result went from around 1200 to 1700! Quake 2 performance also jumped to levels I was expecting from other benchmarks.

Pretty interesting - I had always thought the AGP card was just being treated as a "dumb pipe" sending the Voodoo2s image to the screen, so had not really considered that it could be compromising the performance of everything. The whole system actually seems faster and more responsive though, even just at the desktop. Has anyone else encountered this problem? I think the next thing I'll do is some investigation to try and figure out if something is actually wrong with the TNT2 - I know it has some jumpers on it that let you toggle between various AGP modes, so perhaps one of those is set incorrectly.

From my SS7 Nvidia experiences, I found that the TNT / TNT2 cards, behave differently, than the GeForce cards. With different I had to do all sorts of things to get them to work well.