VOGONS


First post, by Mattix23

User metadata
Rank Newbie
Rank
Newbie

Hi everyone, I'm Matteo and I'm writing from Italy!

Today I wanted to share an experiment that might be useful/interesting for anyone thinking about tinkering with an S3 Virge.

A quick disclaimer first: I was born in 1994 and only got into retrocomputing a few years ago. Since I was just a kid when this hardware came out, I never got to test most of these parts back in the day. Please forgive me if some of this feels obvious or well-trodden ground, for me it is all a brand new discovery.

Some time ago I bought a NOS Sparkle S3 Virge online. It is the 375 model, so a DX variant, but it came equipped with only 2 MB of 50ns EDO RAM.

The attachment Card.png is no longer available

The idea of physically expanding a graphics card's video memory fascinated me quite a bit, so I did some research and ordered four identical memory chips from China for next to nothing. As soon as they arrived, I started my experiment.

For this test, I put together a rig that closely mirrors our old family PC. Back then, though, our machine had no 3D card at all. For years I played exclusively in software rendering, staring at giant pixels and weird, dithered transparency effects, though at a decent 640x480 resolution, because I remember lower resolutions looking way too grainy on my CRT monitor and taking on proportions I just couldn't stand. If my dad had bought an inexpensive S3 Virge for pocket change back then, would it have actually made any difference?

Here are the test system specs:

Sparkle S3 Virge DX 375 PCI
Zida Tomato LX98-AT (Intel 440LX chipset)
Intel Pentium II 300 MHz
128 MB RAM
Creative Sound Blaster 16 Value ISA

Instead of relying on synthetic benchmarks, I picked three games I used to play a ton: Monster Truck Madness, Tomb Raider with the dedicated S3 patch, and Star Wars Jedi Knight: Dark Forces II.

Looking back today, Monster Truck Madness really seems built to run on this kind of hardware, featuring a DirectX 3 engine with very basic effects. Indeed, out of the three, it was the most playable, even though I honestly expected a bit more. You have probably seen Tomb Raider tested a thousand times, yet it remains fascinating to see how a game that received an official S3 patch becomes practically unplayable if you actually want an improvement in visual quality! With only 2 MB of memory, Dark Forces II barely produces any frames at all!

Then came the most exciting and eagerly awaited moment: installing the four extra memory chips into the empty sockets. Even though this was my very first time ever doing something like this, everything went incredibly smoothly on the first try. The card booted up without a hitch and Windows immediately recognized all 4 MB.

The attachment 4MB.png is no longer available

Full of enthusiasm, I launched the games again. The visual improvements are noticeable right away: shapes are sharper, several rendering artifacts disappear, and Lara's famous shadow finally shows up in Tomb Raider. Even Dark Forces II starts pushing out a few more frames. Yet, we still aren't quite there. Let's be honest, I would never have sat down to beat Tomb Raider at that framerate, and in Star Wars it would have been impossible to complete even the first level. Monster Truck Madness, on the other hand, was much more enjoyable.

So I decided to play my final card: overclocking. From reading this very forum, I learned that Made in Korea chips like mine tend to handle higher frequencies than Taiwanese ones, even though my 50ns modules were clearly going to be a bottleneck. And that is exactly what happened. PowerStrip showed a stock clock of 51 MHz, and I managed to push it up to 66 MHz. Those aren't earth-shattering numbers, sure, but it is still a 30% increase, and you can feel every bit of it.

The attachment OC.png is no longer available

The situation changes radically. Monster Truck Madness, already decent before, finally becomes a pleasure to play. Tomb Raider, while not shining in terms of framerate, becomes enjoyable considering the inherent slowness of its engine and Lara's movements. Star Wars takes a massive leap forward compared to the starting point, though it remains too heavy to be truly playable.

Here are some Performance graphs I’ve created to simplify things:

Average Framerate

The attachment FPS1.png is no longer available

Min and Max Framerate

The attachment FPS2.png is no longer available

In the end, this experiment let me experience firsthand a truth we all know, yet I needed to feel on my own skin: the S3 Virge is effectively a 3D decelerator, even when paired with a capable, next-generation CPU like the Pentium II. With software rendering the graphics were rough, sure, but the games were smooth and fun. The Virge chipset strips away that playability without even rewarding you with enough of a visual upgrade to justify the sacrifice.

Despite everything, I walk away with a fantastic experience, full of curiosity and pure fun. Of course, if I had spent my savings on an S3 Virge and a memory expansion back in the nineties only to get these results, I probably would have been a bit disappointed. Or maybe, who knows, I would have made myself like it anyway, after all, there were real results! Too bad about the overclocking. I will forever be left wondering how far the chip itself could have actually been pushed.

For anyone wanting to see the tests live, I uploaded a video on my YouTube channel.

https://youtu.be/gB_cnPzXHP4

I am not a YouTuber and I have no intention of becoming one: I have just accumulated quite a bit of boxed hardware from the 90s to the mid-2000s, and I am trying to give some purpose to all this junk.

What about you guys? Did any of you actually own an S3 Virge back in the day as your main 3D card? How did you feel about it at the time?

Reply 1 of 16, by cloverskull

User metadata
Rank Member
Rank
Member

Cool experiment! I enjoyed your video very much. Nicely done.

Reply 2 of 16, by BinaryDemon

User metadata
Rank Oldbie
Rank
Oldbie

Most of my hate for the Virge comes from my experience with Descent2. I sorta wish you had tested that. Not sure if it was me paring it with a Cyrix P-200 or it was equally bad in every configuration but honestly I stuck with software renderer until I could afford and upgrade to Rendition Verite 2100, and that experience was amazing.

Reply 3 of 16, by kixs

User metadata
Rank l33t
Rank
l33t

Maybe try with lower resolutions. Try the same resolution as software mode and compare frame rates and graphics quality.

Visit my AmiBay items for sale (updated: 2026-03-23). I also take requests 😉
https://www.amibay.com/members/kixs.977/#sales-threads

Reply 4 of 16, by Mattix23

User metadata
Rank Newbie
Rank
Newbie
cloverskull wrote on Yesterday, 20:07:

Cool experiment! I enjoyed your video very much. Nicely done.

Thanks a lot! I'm glad you enjoyed it ☺️☺️☺️

Reply 5 of 16, by Mattix23

User metadata
Rank Newbie
Rank
Newbie
BinaryDemon wrote on Yesterday, 20:27:

Most of my hate for the Virge comes from my experience with Descent2. I sorta wish you had tested that. Not sure if it was me paring it with a Cyrix P-200 or it was equally bad in every configuration but honestly I stuck with software renderer until I could afford and upgrade to Rendition Verite 2100, and that experience was amazing.

Would you believe me if I told you that, while I’m well aware of Descent 2 it’s quite renowned within our vintage community, I’ve actually never played it myself?
Maybe now’s the time 😁. I haven’t dismantled the build yet, so let’s see if I can manage to run a few tests over the next few days.

Reply 6 of 16, by Mattix23

User metadata
Rank Newbie
Rank
Newbie
kixs wrote on Yesterday, 20:39:

Maybe try with lower resolutions. Try the same resolution as software mode and compare frame rates and graphics quality.

For sure lowering the resolutions help. Is well know that Virge gives the best restults ideally at 512. But honestly I really hate that resolution 😩

Reply 7 of 16, by kixs

User metadata
Rank l33t
Rank
l33t

I would go even with lower resolution and check the graphics compared to software mode. Bilinear filtering should improve graphics even at lower resolutions while keeping the FPS playable. You can't expect miracles with a low end 3D card. I'd do it myself but have no time 🙁

People with lower end cards usually also had smaller monitors. So 15" CRT would look just fine even with lower resolution.

It was called 3D decelerator especially as everyone was comparing it against a Voodoo card at 640x480 resolution. It wasn't the case with just S3 Virge but also with ATI Rage II and others...

Visit my AmiBay items for sale (updated: 2026-03-23). I also take requests 😉
https://www.amibay.com/members/kixs.977/#sales-threads

Reply 8 of 16, by marxveix

User metadata
Rank Oldbie
Rank
Oldbie
kixs wrote on Yesterday, 22:38:

It wasn't the case with just S3 Virge but also with ATI Rage II and others...

Original ATi Rage1 (3D Rage) that was more like S3 Virge (bit faulty 3d).

Rage II -> Rage2+DVD ->Rage2C all not so good, but Rage2C AGP is fastest of the slow ones and better to have 4MB at least.
2MB can use 3D,but only up to 400x300 usually and to have playable framerate do not go over 512 at early Rage2s to get fps.

Best ATi Rage3 drivers for 3DCIF / Direct3D / OpenGL / GLUT / DVD : ATi RagePro drivers and software
33+MiniGL / OpenGL Win 9x dll files for all ATi-Rage3 based cards : ATi RagePro OpenGL files

Reply 9 of 16, by kixs

User metadata
Rank l33t
Rank
l33t

It's all so long time ago I'm beginning to forget how it was...

In that time I had Rage II and also tested S3 Virge - no idea what model. I used 3D WinBench 97 or 98. I know it run as a slideshow 🤣 But I never tested in real games at low resolution. At least just for the fun of it. But we all knew that only 3dfx was the way to go for games.

Now I can't even remember on what card did I play Quake II in 1998/99. It might be Nvidia TNT as the card was 2D & 3D.

Visit my AmiBay items for sale (updated: 2026-03-23). I also take requests 😉
https://www.amibay.com/members/kixs.977/#sales-threads

Reply 10 of 16, by The Serpent Rider

User metadata
Rank l33t++
Rank
l33t++

Bilinear filtering is one of the weakest points of S3 Virge. You can gain quite a lot by disabling it where it's possible.

I must be some kind of standard: the anonymous gangbanger of the 21st century.

Reply 11 of 16, by the3dfxdude

User metadata
Rank Oldbie
Rank
Oldbie

As you may know, when we had these cards new in the '96-'97 time frame, Pentium 2's while sort of contemporary, were super expensive. These cards were in slower machines than the P2 for almost everyone, and played with at lower resolutions. S3 was one of the most popular makers of video chips, and sold lots very cheap, and there weren't really any other brands that had this level in the market, so they briefly enjoyed the most popularity for games, and availability. I had these Virge cards and so I know what you are talking about at time went on. When the Voodoo card came out, the Virge was over as a 3D option, but did last a while longer and the primary 2D card. This is why it got the bad rep. Clock speeds increased, resolutions and framerate got higher with the Voodoo, and people could directly compare things like the Voodoo and the Virge in the same system, and saw it really didn't scale as clock speeds got faster. Because lots of people had this card, and never used the Virge unless it was before the Voodoo, the comparisons just poured out as the decelerator, because it couldn't do what the companion Voodoo cards could do.

So as people said, it's supposed to run in slower systems at a lower resolution where it is competitive to software rendering.

For the games back in the day, Tomb Raider and certainly Monster Truck, these games likely would have been enjoyed by people with the Virge. JK2 came later, and I think while it was playable, I remember trying both and preferring software mode on this case.

Reply 12 of 16, by Joseph_Joestar

User metadata
Rank l33t++
Rank
l33t++

Another thing that people often forget is how much better the Virge's S3D mode looked compared to software rendering. Specifically, Tomb Raider has very noticeable color banding when using software rendering, since it runs with 256 colors only (8-bit).

file.php?id=112144&mode=view
file.php?id=112145&mode=view
file.php?id=112146&mode=view

In Virge's S3D mode, Tomb Raider runs with 65536 colors (16-bit) which looks much nicer. And as others have said, if you compare the Virge DX to software rendering on a contemporary Pentium 166 MHz CPU, the frame rates in Tomb Raider will be roughly the same. Lastly, not all Virge cards were made equal. Models from reputable manufacturers such as Diamond or Hercules were often clocked higher than "no name" Virge cards, which could result in a measurable performance boost.

My retro stuff

Reply 13 of 16, by Putas

User metadata
Rank Oldbie
Rank
Oldbie

Nicely written, Matteo. I also enjoy squeezing something out of the low end, it feels much more rewarding.
Always overclock your Virges. The low frame rates combined with vsync can yield disproportionate gains.

Reply 14 of 16, by Ozzuneoj

User metadata
Rank l33t
Rank
l33t

Nice write up!

Still, I would say that a Pentium II 300 with a Virge DX is an example of a huge video card bottleneck. The PII 300 was released as the top end consumer CPU from Intel around six months after the Virge DX with an MSRP of around $1800 just for the CPU (so says the internet). The insane pace at which computers were improving back then makes this kind of crazy to imagine since they were only six months apart, but if someone put a Virge DX in a brand new PII 300 system in 1997, it would have been like putting a GT 740 into a modern PC that has a modern CPU + iGPU. The CPU (and in this case the iGPU) is just much more capable, so the video card is a decelerator.

Mismatches like this probably did happen somewhat frequently back in the day though because people just didn't know or care, and the cards were ending up in PCs for years. This is likely the origin of most of the "3D Decelerator" stuff about the Virge.

A much more suitable pairing for a Virge DX would have been something in the realm of a 133Mhz Pentium or something slower. In that case, games would only really be playable at 320x200 or maybe 400x300 in software mode, and at those resolutions the Virge DX can handle quite a bit more and provide a smoother experience than software.

Now for some blitting from the back buffer.

Reply 15 of 16, by mkarcher

User metadata
Rank l33t
Rank
l33t
Mattix23 wrote on Yesterday, 20:47:

For sure lowering the resolutions help. Is well know that Virge gives the best restults ideally at 512. But honestly I really hate that resolution 😩

In my oppinion, one of the key points of the ViRGE's failure in the PC market is missing software using 512*384. 640*480 is definitely too much for a classic ViRGE (86c325) at its stock clock (50MHz), but for example in Terminal Velocity that is the only resolution supported with 3D acceleration. Furthermore, Terminal Velocity only uses 2MB of video RAM, in which 1200KiB are already used by the double-buffered 16-bit frame-buffer, so only 848K remain for textures. If I remember correctly, this caused the textures in the ViRGE edition to have a lower resolution than the optional "high resolution" texture you got in the paid version. My hunch is that porting PC games that used SW rendering at 320*200 or 320*240 to 640*400/640*480 was way easier than to 512*384, just because bitmap assets could be re-used at 2:1 scaling, which is easy to do.

Having a ViRGE/DX, you can get usable frame-rates at 66MHz at 640*480, at least if you ditch bilinear filtering and/or perspective correction. On the original ViRGE, you would have to disable both to get kind-of usable rates at 640*480. The /DX both increased the maximum specified clock and the efficiency of perspective correction, reducing the impact of these "high-end" features.
Originally (e.g. hardware-accelerated Quake), the driver unconditionally enabled both perspective correction and bilinear filtering, and people rightfully compared it to the Quake software renderer. Now, Quake is actually a special case: It had one of the best optimized software renderer for Intel processors available at that time, which included perspective correction and yielded usable frame rates on Pentium II class computers. So to make the hardware accelerated 3D be "better" than software rendering, you may not disable perspective correction (which would look awfully distorted compared to the perspective corrected software renderer), and you would want to excel at something, which is bilinear filtering (which the Quake software renderer did not have). Comparing the "perfect" sw renderer to a mediocre hardware renderer, with the hardware renderer operating at a higher target quality, caused the unsurprising result of the hardware being slower.

Reply 16 of 16, by leileilol

User metadata
Rank l33t++
Rank
l33t++

Overclocking won't overcome the ViRGE's awful depth precision and alpha modulation for RGBA4444 textures however (other than the obvious lack of blending functionality)

FUCK "AI". It is a tool of fascism. We do not need it. We do not use it.