GPU clocks are highly likely to be slower on the Hercules vs the MX400. Then also FX cards can be all over the place with memory bus width and bandwidth.

PCI cards suffer heavy penalty in games with T&L support, especially pre-GeForce FX cards. Quake 2 have transformation part supported by OpenGL.

Also here are my tests on PMMX 233: Intel 430HX L2 256Kb VS Intel 430HX L2 512Kb VS Intel 430TX L2 512Kb - Motherboards comparison

BitWrangler wrote on 2022-03-13, 16:22:

GPU clocks are highly likely to be slower on the Hercules vs the MX400. Then also FX cards can be all over the place with memory bus width and bandwidth.

So you're saying the cards should perform the same?

I don't know about the hercules card, but the Leadtek MX400 runs at 175 core and 166MHz ram. It's a 128 bit card, and judging by the number of memory chips so is the Hercules.

The FX5500 PCI runs at 250MHz core and 500MHz (250x2) for the ram witch runs on a 128bit bus (verifyed with GPU-Z). The chaintech FX 5200 runs at 275MHz core (gpu-z reports 278) and 500MHz (250x2) ram, also on 128 bit bus. I don't keep any 64 bit cards, except for maybe TNT2 M64's, S3 Savage cards or low profile FX5700LE. All the 64 bit 5200's and radeon 9000/9200 card's I've gotten have either been used for parts (caps, memory chips, SMD components) then recycled or experimented on (managed to turn a 5200 and a 9200 from 64 to 128 bit by populating the empty ram pads and supporting SMD components).

[EDIT] I'm pretty sure the 3D Prophet II MX PCI runs at 183Mhz core.

The Serpent Rider wrote on 2022-03-13, 16:39:

PCI cards suffer heavy penalty in games with T&L support, especially pre-GeForce FX cards. Quake 2 have transformation part supported by OpenGL.

Also here are my tests on PMMX 233: Intel 430HX L2 256Kb VS Intel 430HX L2 512Kb VS Intel 430TX L2 512Kb - Motherboards comparison

Thanks for the link, reading your topic now.

Socket3 wrote on 2022-03-13, 16:41:

[EDIT] I'm pretty sure the 3D Prophet II MX PCI runs at 183Mhz core.

Not that it makes much difference, but the AGP version of that card I had back then ran at 175MHz core, 183MHz 128-bit SDRAM. The driver came with Coolbits enabled but the GPU wasn't overclocked by default.
I've never seen the PCI version though, but I'm surprised if it's clocked faster.

[LONG POST INCOMING]

Ok so I managed to find the time to do a clean install of win9x on one of my machines and re-ran some benchmarks. Here is the hardware configuration:

- AMD K6-2+ 500Mhz @ 500MHz
- Aopen AX59 PRO VIA MVP3
- 128MB SDram PC133 @ 100MHz CL2
- 20GB Seagate HDD
- Yamaha DX-XG (YMF722) sound card

And the software:

- Windows 98 Second Edition, no service packs, no patches
- Voodoo 3 Drivers ver 1.07
- DirectX 8.1
- Quake 2 3.20
- 3D Mark 2000
- Via 4in1 4.43

The PC:

AGP Voodoo 3 3000:

PCI Voodoo 3 3000:

As you can see, the PCI voodoo 3 is the SGRAM version, while the AGP card uses SDRAM. Both run at 166/166Mhz.

And here are my results. First the AGP card:

And now the PCI card:

As you can see, this time around the PCI card leads by a very small margin, probably due to the fact that it's using faster SGRAM.

V3 3000 SDRAM AGP:
-Quake 2 800x600: 57.2 FPS
- 3D Mark 2000 @ 800x600 16 bit color: 1622 PTS

V3 3000 SGRAM PCI:
-Quake 2 800x600: 58.6 FPS
- 3D Mark 2000 @ 800x600 16 bit color: 1649 PTS

I don't know about 3dfx PCI cards I never compared them with the AGP versions in the time correct systems but older cards there was quite a difference in speed when testing AGP or PCI versions, like the Rage Pro AGP vs PCI. I don't know what and where is the limitation, maybe the chip and card was configured in a different way (while at the same clocks) but the difference sometimes was too much to accept. If the AGP version is more or less "enough" fast for time correct or later games, the PCI version isn't mostly anywhere at the point even the Rage IIc AGP seems like more or less similar.

About the modern latest PCI cards running with a IC bridge, from my test the bridge choice itself can make a difference depending also on the system is tested on and the speed of the whole config. The final speed is quite variable at the end and mostly limited.

386SX wrote on 2022-03-23, 17:36:

I don't know about 3dfx PCI cards I never compared them with the AGP versions in the time correct systems but older cards there was quite a difference in speed when testing AGP or PCI versions, like the Rage Pro AGP vs PCI. I don't know what and where is the limitation, maybe the chip and card was configured in a different way (while at the same clocks) but the difference sometimes was too much to accept. If the AGP version is more or less "enough" fast for time correct or later games, the PCI version isn't mostly anywhere at the point even the Rage IIc AGP seems like more or less similar.

About the modern latest PCI cards running with a IC bridge, from my test the bridge choice itself can make a difference depending also on the system is tested on and the speed of the whole config. The final speed is quite variable at the end and mostly limited.

Thanks for the input. My earlier testing with the Geforce 2 MX 400's and FX 5200/5500 AGP and PCI back up your experience. I dug up a Inno3D FX5500 to run against my PCI 5500 and I plan to do the same tests with the newer cards. I plan to use a much faster system then the K6 - possibily a socket A or socket 478 build with a fresh install of windows 98. I suspect the voodoo 3 / K6-2+ combo is not fast enough to saturate the PCI bus, and together with the fact that the voodoo 3 does not use AGP features, my results make sense.

But I think there will be a noticeable difference in performance between the two FX 5500 cards. I think they might be fast enough to saturate the PCI bus.

What I'm trying to figure out is why my Compaq D500 runs so poorly with the FX 5500 PCI. 3D Mark 2001 results are on par with what a FX5200/FX5500 would score on a pentium III platform, but some games run poorly. Mechwarrior 3 will occasionally slow down for example - to what feels like 20fps. Ground control is completely unplayable. In contrast Dungeon Keeper 2 and Unreal Tournament 99 run great. I don't know if there's an issue with the PC itself (considering it was designed for office work) or the PCI bus. BIOS is very rudimentary. No CPU latency, no CPU to PCI settings, no memory speed settings, nothing. It might just be slow.

I'll post a comparison between the FX 5500 cards as soon as I find the time to set up a system and run benchmarks.

D500 has 100/400Mhz FSB Willamette CPUs with SDRAM and you're asking why it only runs at PIII speeds... well.... because it does.... unless you've got the hotrod 2.0 in there, they barely pull away from the top end P3s.

Socket3 wrote on 2022-03-23, 17:58:

386SX wrote on 2022-03-23, 17:36:

I don't know about 3dfx PCI cards I never compared them with the AGP versions in the time correct systems but older cards there was quite a difference in speed when testing AGP or PCI versions, like the Rage Pro AGP vs PCI. I don't know what and where is the limitation, maybe the chip and card was configured in a different way (while at the same clocks) but the difference sometimes was too much to accept. If the AGP version is more or less "enough" fast for time correct or later games, the PCI version isn't mostly anywhere at the point even the Rage IIc AGP seems like more or less similar.

About the modern latest PCI cards running with a IC bridge, from my test the bridge choice itself can make a difference depending also on the system is tested on and the speed of the whole config. The final speed is quite variable at the end and mostly limited.

Thanks for the input.

I wonder if the real benefit of the AGP chips were to be designed specifically for that in mind, with the bus higher clock and reserved connection instead of the even good PCI usage of the Riva128 using the system memory too or the (as seen above) good PCI usage of the 3dfx chips I suppose more for their internal design (that at the end was born in the PCI times and not changed that much later) beside a lack of the AGP memory usage that would have not changed much if not anything at all.
I suppose it's not about the PCI bandwidth (that would make all the later PCI bridged MUCH faster video cards like the Geforce GT520/610 totally useless if the PCI bus was already limiting in those times..) but more about the time correct combination of hw they were oriented to and how good the drivers were; maybe some of the early AGP chips used in the PCI version worked in some "compatibility design" different than the 3dfx ones that seems like designed for it. PCI bus later has driven much more demanding tasks on bridged GPUs (even if still far from the AGP/PCI-EX versions but as said IC bridges <> chipset bridges combination/latency imho make the whole logic even more limited but still capable of running complex games; when I tested the Geforce GT610 PCI I read a lot about different IC bridge chips features and it seems like they were quite complex IC trying to keep both speed and latency at best and probably not really oriented for GPUs at first).

386SX wrote on 2022-03-23, 20:15:

PCI bus later has driven much more demanding tasks on bridged GPUs (even if still far from the AGP/PCI-EX versions but as said IC bridges <> chipset bridges combination/latency imho make the whole logic even more limited but still capable of running complex games; when I tested the Geforce GT610 PCI I read a lot about different IC bridge chips features and it seems like they were quite complex IC trying to keep both speed and latency at best and probably not really oriented for GPUs at first).

Another reason to never touch later gen PCI cards - particularly cards that were designed to run on PCI Express like the 520 and 610. Add to that the fact that neither of those cards have win9x drivers (the OS where I do most of my retro gaming). I don't think the same is true for PCI-E to AGP bridge chips. I've come across x1950, 4650 and 3850 AGP cards and I have to say they performed impeccably - despite all of them having a bridge chip - but again, none of those have win9x drivers, and for windows XP it's much easier to build a PC with PCI express.

If performance is seriously hampered on the newest PCI cards like the ones you mentioned, it means they have near zero retro gaming value. They're not fast cards to begin with. They were probably meant for use in servers or industrial computers.

BitWrangler wrote on 2022-03-23, 19:44:

D500 has 100/400Mhz FSB Willamette CPUs with SDRAM and you're asking why it only runs at PIII speeds... well.... because it does.... unless you've got the hotrod 2.0 in there, they barely pull away from the top end P3s.

The Deskpro D500 comes in either socket 370 or socket 478 flavors. The one I have is almost identical to the Deskpro EN SFF and came with a 1300MHz Tualatin celeron but performs like a 500Mhz mendocino in some games. Read the first two posts please. I'm not expecting miracles - It should be able to run ground control (a game from 2000) flawlessly, especially with the 1.4Ghz Tualatin I put in it - but it's not. The computer I built around an Abit ST6 runs laps around it in some games, while in others it performs roughly the same. Mechwarrior 3 can run on a potato, it should run exceptionally on the D500. In fact the time I completed the game was on a 400MHz K6-2 with on board Trident Blade 3D graphics and the game was very playable @ 640x480.

download/file.php?mode=view&id=133418

Socket3 wrote on 2022-03-23, 21:55:

Another reason to never touch later gen PCI cards - particularly cards that were designed to run on PCI Express like the 520 and 610. Add to that the fact that neither of those cards have win9x drivers (the OS where I do most of my retro gaming). I don't think the same is true for PCI-E to AGP bridge chips. I've come across x1950, 4650 and 3850 AGP cards and I have to say they performed impeccably - despite all of them having a bridge chip - but again, none of those have win9x drivers, and for windows XP it's much easier to build a PC with PCI express.

I suppose the last PCI video cards for PCI-EX GPUs were probably oriented to a very thin market but I think it's not bad to have such alternative solutions even for those OEM old boards having modern CPU on some proprietary layout with PCI only bus or maybe mini-itx oriented. Of course the final speed is not really gaming oriented but for old games they can still work enough fast. For example I used them on Atom based mini-itx where it benefit a lot from it and the heavy modern GUIs can benefit from the WDDM1.2 or 1.3 features while the cpu speed limitation is another factor to add to the "castle of IC bridges" solution (PCI-EX<>PCI to PCI<>PCI-EX1 and back).
About the AGP bridges I suppose the AGP logic @ 8x speed had more space to keep final speed close to the time correct PCI-EX solution. Also being proprietary IC designed specifically for their own GPUs.