It doesn't matter in which order the drivers are installed. Usually. But apparently in Ydee case the AGP problem is fixed when VIA GART driver is installed last.

The Serpent Rider wrote on 2025-11-04, 16:23:

It doesn't matter in which order the drivers are installed. Usually. But apparently in Ydee case the AGP problem is fixed when VIA GART driver is installed last.

It's difficult to install the network card driver first if it's not listed in the OS device manager at all.
I wouldn't be writing about this with such confidence if I hadn't spent half a day replacing several network cards 20 years ago.
That's why I still remember these subtleties.

In Windows 98, 4-in-1 technology almost completely reconfigures the chipset, as the VIA support implementation in the BIOS\OS build-in drivers turned out to be less than ideal.
At least in the Award BIOS.

I have to admit, I have the onboard LAN, AC97,LPT and COM disabled in my BIOS. So I don't know if the W98 would have any problem without installing 4in1 drivers if I had everything enabled, but I turn off anything I don't use.
I just wanted to write that when installing 4in1 drivers as the first software after installing the OS itself, the AGP texture acceleration was disabled and unavailable. When I installed the 4in1 package after forceware drivers were installed, the acceleration switched on and can be turned off - before the button was inactive.

Of course, I don't know if this will help solve the buggy problem every time, but I haven't had a chance to get any older versions of the BIOS (the oldest version is from 2006), changing the speed of the AGP (8x to 4x) hasn't helped, and other tips on the solution haven't worked either. The procedure described above solved the problem in my case.

Detonator most likely reads the state of the chipset registers; if it doesn't like something, it disables AGP memory.

At the time, Nvidia drivers were actively changing modes due to stability issues with VIA-based motherboards.

It would be interesting to compare Nvidia driver registry keys for different installation methods.
Perhaps we can identify key something like "AGP RAM disabled" or "Capability mode on."

Ydee wrote on 2025-11-04, 15:33:

I don't know if I've come across a one-size-fits-all solution, but I'll describe what worked in my case.
<...>

Thanks for the info. I've tried my best to replicate this on my 4CoreDual-VSTA with GeForce 5700LE. Unfortunately - no dice.

dxdiag does indeed say that AGP Texture Acceleration is enabled (was it ever unavailable on 4CoreDual?), however 3dmark2001 still says that Total AGP Memory is 0 bytes. AGP capabilities tab shows that Aperture Size is 0 bytes.

The system scores ~7300 marks in 3DMark 2001SE (Core E7600, Geforce 5700LE).

Ah well.

shevalier wrote on 2025-11-04, 16:48:

In Windows 98, 4-in-1 technology almost completely reconfigures the chipset, as the VIA support implementation in the BIOS\OS build-in drivers turned out to be less than ideal.
At least in the Award BIOS.

That's fascinating and explains a lot of the weirdness of earlier VIA chipsets. However, besides this AGP issue on later BIOS versions, the PT880 and K8M800 chipset based motherboards are perfectly happy with just Windows INF based installs. I automated viamach.inf, usb2via.inf, viamraid.inf, and viagart.inf installation on 98SE for testing purposes (infinst.exe) and all the device drivers install without any issues. I had to edit in missing registry entries to viagart.inf but that was all I had to modify. I've been able to completely avoid using the 4in1 executable. What chipset did you encounter those issues on?

The Serpent Rider wrote on 2025-11-04, 16:23:

It doesn't matter in which order the drivers are installed. Usually. But apparently in Ydee case the AGP problem is fixed when VIA GART driver is installed last.

It does matter. When researching this a while ago I found that VIA's AGP implementation was always pretty shaky and prone to not installing correctly, leading to the same symptom since the GART driver isn't present or installed correctly. The "correct" method I've read is going through the 4in1 installer's items one at a time, not all at once.

I see some of that happening in this thread too. But it's not the solution to the specific issue with VIA PT8XX and K8XXXX chipsets after the 2005 BIOS update that fundamentally broke GART in 98. Some posters are also encountering Nvidia's soft GART driver, implemented some time after 45.XX. IIRC that will show 32 or 64 MB of AGP memory in 3DMark01 regardless of what the setting is in BIOS, because it's not hardware based GART provided by the chipset. It's better performance sure, but that's also not a real solution to this issue.

shevalier wrote on 2025-11-05, 06:46:

Detonator most likely reads the state of the chipset registers; if it doesn't like something, it disables AGP memory.

I might have ruled that out last year when I compared all the registers between the known good and known bad BIOS versions, all of the chipset and GPU registers are identical.

At the time, Nvidia drivers were actively changing modes due to stability issues with VIA-based motherboards.

It would be interesting to compare Nvidia driver registry keys for different installation methods.
Perhaps we can identify key something like "AGP RAM disabled" or "Capability mode on."

I tried every flag I could find regarding VIA, AGP, or GART in RivaTuner, nothing helped this issue. Those all seemed to be for issues with older chipsets, at least that's what I gleamed from read 2 decade old forum posts. I think trying to install the GPU driver before the chipset and AGP driver could be the cause of some of those issues.

Zoomer wrote on 2025-11-05, 12:30:

dxdiag does indeed say that AGP Texture Acceleration is enabled (was it ever unavailable on 4CoreDual?), however 3dmark2001 still says that Total AGP Memory is 0 bytes. AGP capabilities tab shows that Aperture Size is 0 bytes.

That is something that should be clarified in the OP, dxdiag is not a reliable report if the bug is present or not. The most reliable way I can think of is: if AGP is properly installed and enabled in the Nvidia display settings tab:

And, like in the OP, 3DMark01 reports 0 bytes of AGP memory, then the bug is most likely present. If the AGP memory values are mismatched between the Nvidia display settings and 3DMark01, it might be because it's running a soft GART.

VDNKh wrote on 2025-11-19, 15:22:

That's fascinating and explains a lot of the weirdness of earlier VIA chipsets. However, besides this AGP issue on later BIOS ve […]
Show full quote

shevalier wrote on 2025-11-04, 16:48:

In Windows 98, 4-in-1 technology almost completely reconfigures the chipset, as the VIA support implementation in the BIOS\OS build-in drivers turned out to be less than ideal.
At least in the Award BIOS.

That's fascinating and explains a lot of the weirdness of earlier VIA chipsets. However, besides this AGP issue on later BIOS versions, the PT880 and K8M800 chipset based motherboards are perfectly happy with just Windows INF based installs. I automated viamach.inf, usb2via.inf, viamraid.inf, and viagart.inf installation on 98SE for testing purposes (infinst.exe) and all the device drivers install without any issues. I had to edit in missing registry entries to viagart.inf but that was all I had to modify. I've been able to completely avoid using the 4in1 executable. What chipset did you encounter those issues on?

The Serpent Rider wrote on 2025-11-04, 16:23:

It doesn't matter in which order the drivers are installed. Usually. But apparently in Ydee case the AGP problem is fixed when VIA GART driver is installed last.

It does matter. When researching this a while ago I found that VIA's AGP implementation was always pretty shaky and prone to not installing correctly, leading to the same symptom since the GART driver isn't not present or installed correctly. The "correct" method I've read is going through the 4in1 installer's items one at a time, not all at once.

I see some of that happening in this thread too. But it's not the solution to the specific issue with VIA PT8XX and K8XXXX chipsets after the 2005 BIOS update that fundamentally broke GART in 98. Some posters are also encountering Nvidia's soft GART driver, implemented some time after 45.XX. IIRC that will show 32 or 64 MB of AGP memory in 3DMark01 regardless of what the setting is in BIOS, because that's not hardware based GART provided by the chipset. It's better performance sure, but that's also not a real solution to this issue.

shevalier wrote on 2025-11-05, 06:46:

Detonator most likely reads the state of the chipset registers; if it doesn't like something, it disables AGP memory.

I might have ruled that out last year when I compared all the registers between the known good and known bad BIOS versions, all of the chipset and GPU registers are identical.

At the time, Nvidia drivers were actively changing modes due to stability issues with VIA-based motherboards.

It would be interesting to compare Nvidia driver registry keys for different installation methods.
Perhaps we can identify key something like "AGP RAM disabled" or "Capability mode on."

I tried every flag I could find regarding VIA, AGP, or GART in RivaTuner, nothing helped this issue. Those all seemed to be for issues with older chipsets, at least that's what I gleamed from read 2 decade old forum posts.

Zoomer wrote on 2025-11-05, 12:30:

Ydee wrote on 2025-11-04, 15:33:

I don't know if I've come across a one-size-fits-all solution, but I'll describe what worked in my case.
<...>

dxdiag does indeed say that AGP Texture Acceleration is enabled (was it ever unavailable on 4CoreDual?), however 3dmark2001 still says that Total AGP Memory is 0 bytes. AGP capabilities tab shows that Aperture Size is 0 bytes.

That is something that should be clarified in the OP, dxdiag is not a reliable report if the bug is present or not. The most reliable way I can think of is: if AGP is properly installed and enabled in the Nvidia display settings tab:

Where it says "AGP 4X" and "128 MB"

The attachment agp_nv_display.png is no longer available

And, like in the OP, 3DMark01 reports 0 bytes of AGP memory, then the bug is most likely present. If the AGP memory values are mismatched between the Nvidia display settings and 3DMark01, it might be because it's running a soft GART.

It was definitely a VIA Pentium 3 implementation from ASUS.
Most likely something on the 694 chipset.
Definitely not the MVP3, whose problems have become legendary.

As for the latest chipsets from VIA, there were no particular complaints when they were relevant.
But back then, hardly anyone used them with Windows 98.

You can add the Asus A8V-X to the list of boards with this issue (As well as the Asus P5PE-VM)...

No early BIOS available for it, and you can't cross-flash it because it's a bit unique (the southbridge on it is rare).

I did have partial success with the nVidia driver workaround (driver 81.98 for example gives me 24MB of AGP memory in 3DMark 2001SE).

I don't think this is strictly a VIA issue. I also tested on my Intel 865G platform which also shows 0MB in 3DMark. As a sanity check, my Thunderbird KT133A system shows 128MB with the same driver (45.23, DX 8.1b), and my P4 Northwood (845) shows something like 98MB in 3DMark 2000..

Win98 has no mechanism for GART other than through AGP, so PCIe systems in Win98 with no AGP texture acceleration at all will also lose performance, and is therefore not a solution to this problem.

If I had to speculate, I would say that swapping the AMD AGESA with from the old BIOS to the new would solve the issue (with the compromise of losing support for newer CPUs).

In my estimation, this is not a platform worth pursuing for Windows 98.

In practical sense, this is mostly non-issue for Windows 98.

The Serpent Rider wrote on 2026-02-20, 08:38:

In practical sense, this is mostly non-issue for Windows 98.

It's been ages since I tested that, but I remember there being a pretty substantial hit in both 3DMark 2001 and game performance under Win98 compared to WinXP, while the bug was present. The difference was around 40%, unless I'm misremembering.

I don't have access to that setup anymore, so I can't double check, but maybe someone else can provide a comparison.

Well, depends which games: if it's something more complex than Quake 3, then Windows 2k/XP is a better choice to play it anyway. And if it drops from 200+ fps to 120+ fps in older games, then it's also not a big practical issue.

The Serpent Rider wrote on 2026-02-20, 09:06:

Well, depends which games: if it's something more complex than Quake 3, then Windows 2k/XP is a better choice to play it anyway. And if it drops from 200+ fps to 120+ fps in older games, then it's also not a big practical issue.

The drop is significant, especially in 1% / 0.1% lows, there are a lot of visible stutters even if the average framerate is "very good".
Of course, you're right, most of these games would work just fine on WinXP, but some of us just want to see Win98 games flying at 500+ FPS. 😁

bloodem wrote on 2026-02-20, 09:19:

The drop is significant, especially in 1% / 0.1% lows, there are a lot of visible stutters even if the average framerate is "very good".
Of course, you're right, most of these games would work just fine on WinXP, but some of us just want to see Win98 games flying at 500+ FPS. 😁

The question then becomes if this manifests with high-end PCIe GPUs on 98se. There's no way 98se is doing any sort of texture acceleration with a PCIe GPU. So if you take something like a 6800GT, and test it on systems of similar spec, one PCIe, one AGP, are you going to see a great deal of a difference?

...or is the on-board 256mb of RAM going to compensate for the lack of texture acceleration.

How do you recommend I test this?

I was able to obtain an Asus A8V today. Luckily, my board (Revision 2.01) comes with the VIA VT8237R Plus southbridge so I don't have to mess with SATA hullabaloo in limiting drives to SATA 150 and being confined to 2015 drives with Sandforce controllers.

The board came with BIOS 0219 from November 2005, so I went with it... This BIOS version does not allow you to use the SATA without loading the option ROM for it. I would assume that they changed this in later BIOS versions (like with the A8V-X which even allows you to select AHCI with the VT8251)... This is not a big deal for me.

Most importantly, the A8V is NOT in fact affected (thus confirming @bloodem's post from 2022 (though in this case I am using the actual hardware).

The AGP aperture is set to 256MB in the BIOS, and with DirectX 9.0C and GeForce 45.23 drivers, it is showing 128mb of texture memory in 3DMark. I didn't even reinstall Windows, this is with the installation from the A8V-X.

With the 3DMark benchmark, there is a significant gain in performance:

A8V-X

A8V

I will test with later BIOS versions to see which (if any) break it, and I will report back.

EDIT (02/22/2026 19:42 EST): I can confirm proper operation with this board even with the last BIOS 0229 from May 2007. This will be a very good candidate for my dual XP/98 build indeed.

Next I will test the AsRock 775i65G (revision 2.04). It, along with the Asus P5PE-VM are the only two other AGP boards that can give AGP dual core AMD K8 (socket 754/939) a run for its money. AGP is broken the same way with the P5PE-VM, but we'll see if AsRock is any better. A 45nm Pentium E5800 will crush any K8 of that era.

I tested my AsRock 775i65G (revision 2.04, BIOS L3.22), it is not affected by the bug. With the AGP aperture set to 256MB, 3DMark2001SE reports 128MB of AGP memory available (at a 64MB aperture, it reported only 32MB).

I also ran the benchmark, and with a Pentium 4 HT 2.8E, I scored 12101. Soon I will test with a 45nm E5800 and see by how much it crushes the K8.

So, there's another viable platform for a high-end 98 AGP machine.

The E5800 Intel combo trounces the AMD 939 equivalent. Besides possessing a huge performance-per-clock improvement over the P4, it also has a good 500mhz-1ghz processor clock advantage over socket 939 AMD. With the same videocard and driver, 3DMark2001SE gains around 2000 points:

E5800 (3.2Ghz), GeForce FX5800, Forceware 45.23:

And here's something a little extra. With a 6800GT, the test results almost double:

E5800 (3.2Ghz), GeForce 6800GT, Forceware 77.72:

A 6800GS scores approximately 1000 points less:

These tests were done with DirectX 8.1b. With the 6800GT, AGP memory showed up as 120mb instead of 128mb. Installing DirectX 9 had no effect on AGP memory, and it functioned properly (I tested this with the A8V only, though I'm certain I would have the same results with the 775i65G).

Now, we compare this with a PCIe system. The system has a 3.6Ghz (oc'd) Wolfdale Core2 with 2GB of DDR2 (the AGP systems were run with only 512mb of DDR). The GPU is a PCIe Quadro FX3450 (poor man's 6800GT).

Wolfdale @3.6Ghz, 2GB DDR2, Quadro FX3450, Forceware 77.72:

The PCIe blows the fastest AGP system out of the water. Although it is only 400mhz faster, the increased speed of other subsystems pushes it far ahead.

Now, @bloodem has stated "The drop is significant, especially in 1% / 0.1% lows, there are a lot of visible stutters even if the average framerate is 'very good'".

The reason for this, at least in theory, is that there is no mechanism in Windows 98se for AGP texturing with PCIe. For later versions of Windows, DXDIAG will show "AGP Texturing" as being 'enabled', so they do employ some compatibility later for it (that means to say, even on a modern Windows 11 system with modern hardware, decades removed from AGP, you still see it). But with 98se, you are stuck with only the memory on the graphics card, whereas proper AGP texturing would then have that amount of RAM from the system shown as being available for the card in 3DMark 2001SE AGP info for more processing.

Does anyone have any suggestion on how I can gather empiricisms to prove that? Is it playing games at resolutions higher than 1024x768 that would exhaust the 256mb of RAM on these cards or is it enabling features like anti-aliasing or anisotropic filtering?

I have a PCIe 7900GTX as well I could test with, though I would never use it on a permanent basis because of the 82.69 drivers which in my opinion are not completely suitable.

There are tools to measure VRAM usage. RivaTuner monitoring for example.

The Serpent Rider wrote on 2026-02-24, 18:13:

There are tools to measure VRAM usage. RivaTuner monitoring for example.

I'm going to install something very taxing like Flatout 2, enable all the visual bells and whistles, and race a few laps. If I don't see any stuttering, then I'm going to conclude that AGP+Aperture is not in fact superior to PCIe without an aperture.

I played a couple of laps of the Desert Oil Field map on Flatout 2. Everything ran fine, and contrary to @bloodem's claim, there was zero microstuttering. The resolution I played at was 1024x768. After 1 lap, I proceeded to set the internal game settings to maximum anisotropic filtering and 4x antialiasing. It looked gorgeous and ran silky smooth without any issues whatsoever.

I also ran 3DMark on a couple more systems (Quadro FX3450 PCIe).

1)ASRock M3A770DE, Athlon II 245e (Regor, 2.9Ghz):

2)MSI K9VGM-V Athlon 64 x2 5000+ (Brisbane, 2.6Ghz)

The VIA PCIe platform's lack of performance has nothing to do with VIA per se, but rather with the paltry cache of the Brisbane (512KB per core, so only 512KB in Windows 98), which is the maximum supported CPU apart from something else obscenely expensive. Regor on my RX780 (AMD 770 chipset) fares a little better with 1MB of cache per core. But it seems 3DMark really likes L2 cache, because my Intel Woldfale (6MB total, 3MB per core), which has 6 times as much cache as the Brisbane and 3 times as much as the Regor in Windows 98, trounces everything. The Regor, as expected, is on par with the E5800 (also has 1MB of cache per core but is 300mhz slower). Despite the Regor platform having quadruple the theoretical memory bandwidth of the E5800 (DDR1 @ "400mhz" vs. DDR3 @ "1600Mhz", or PC-3200 vs PC3-12800), it seems not to affect 3DMark. Cache is king.

In conclusion, PCIe is a perfectly good platform for Windows 98 despite the lack of AGP texturing. It takes a bit of fiddling to get it working right (I am using an ASRock M3A770DE with an Athlon II 245e), but Oerg866's Windows 98 quickinstall was greatly helpful (Windows 98 and PCIe has made strides... Everything is functioning with zero unknown devices or resource conflicts in device manager... I am even using SATA in AHCI mode within Windows 98). Don't go looking for expensive and difficult to find AGP systems, it's not worth your time or money.

Some of you must be wondering how much can you gain over the 6800GT with a 7800GTX (best case scenario -- that means to say -- using a PCIe system while retaining legacy components like floppy, ide, etc...):

Wolfdale @3.6Ghz, 2GB DDR2, GeForce 7800GTX, Forceware 82.69:

Definitely go PCIe and not AGP.