Personnally, my P133 came with a S3 Trio64V2/DX, it just works by itself, though I think on some S3s you can add additional VRAM chips to increase the card's VRAM.

DOS only: VESA compatiblity is key, S3 beats Matrox easily (although early nVidia and 3DFx are arguably even better).

Windows: RAMDAC and analog filter support for clarity at higher resolution is key. S3 depends on card vendor for filters and RAMDACs generally so-so, Matrox wins easily here.

Honestly, the performance difference between the two (at least in DOS games) is maybe 5%. I'd go with the S3 for compatibility purposes even though it's slightly slower. I can tell you that the Millennium II has a different text font in DOS and some games just act weird when changing graphics modes. They end up working but there's something about the way the card handles graphic mode changes that's off-putting. And sometimes the graphic modes are way out of adjustment so you have to dig into your monitor settings to re-center. It's much more at home in Win9x and early Windows gaming IMO.

Honestly, the performance difference between the two (at least in DOS games) is maybe 5%

Actually there's a huge difference between them (in favor of Matrox), but to tap that potential, much powerful system is required. And yes, Matrox Millentium II is quirky for DOS.

The answer is simple: It depends.

GTA1's VRAM recommendation's more about using hi-res 16/24bpp modes, it'll be fine.

The Serpent Rider wrote on 2020-01-29, 12:17:

Honestly, the performance difference between the two (at least in DOS games) is maybe 5%

Actually there's a huge difference between them (in favor of Matrox), but to tap that potential, much powerful system is required. And yes, Matrox Millentium II is quirky for DOS.

A huge difference in DOS? Unless a game specifically supports Matrox (and I believe none do) the card is just acting as a dumb frame buffer, bottlenecked by the PCI bus. No matter what CPU you throw at it, that's not going to change.

the card is just acting as a dumb frame buffer, bottlenecked by the PCI bus

Yes, dumb faster frame buffer. PCI bus limit isn't a problem for high resolutions. Let's be real here: S3TRIO64V+ isn't a pinnacle of 2D performance.

No matter what CPU you throw at it, that's not going to change.

And Vlask benchmark results are just a myth.

dionb wrote on 2020-01-29, 16:49:

The Serpent Rider wrote on 2020-01-29, 12:17:

Honestly, the performance difference between the two (at least in DOS games) is maybe 5%

Actually there's a huge difference between them (in favor of Matrox), but to tap that potential, much powerful system is required. And yes, Matrox Millentium II is quirky for DOS.

A huge difference in DOS? Unless a game specifically supports Matrox (and I believe none do) the card is just acting as a dumb frame buffer, bottlenecked by the PCI bus. No matter what CPU you throw at it, that's not going to change.

I measured a 20% difference between the Millennium II and the slowest S3 card in VGA on a 933Mhz P3. That went up to 30% in SVGA games. It does get more significant the faster the cpu, but I still wouldn't call it "huge". See: Re: 10-Way PCI Graphics Roundup - DOS
But on even a very fast system that was contemporary with DOS games, that difference is much lower. On a 200Mhz Pentium, I measured 4% in VGA and 10% in SVGA.

It's possible those differences become greater when you throw FASTVID/MTRRLFBE into the mix. But that gets into pretty rare territory. The OP is talking about a P1 MMX, and I doubt many people build fast Pentium 3s or Pentium 4s specific as DOS gaming systems.

edit: Even Vlask's benchmarks show a 30% difference between Mill2 PCI and S3Trio64 PCI (92 fps vs 71 fps). So I guess if 30% is the definition of huge...

So I guess if 30% is the definition of huge..

Vlask setup isn't the fastest one, just convenient for ISA cards.

Even Vlask's benchmarks show a 30% difference between Mill2 PCI and S3Trio64 PCI (92 fps vs 71 fps).

67 vs 92 actually. And your experience might get even worse with a "noname" S3 Trio64V+.

The early matrox cards have bios-patches and less problems than the later ones.

dionb wrote on 2020-01-29, 16:49:

A huge difference in DOS? Unless a game specifically supports Matrox (and I believe none do) the card is just acting as a dumb frame buffer, bottlenecked by the PCI bus. No matter what CPU you throw at it, that's not going to change.

Yes the bus (or better say host interface speed) is the most important factor for DOS performance.

However, there is still enough a space for differentiating chips in DOS.

Such as many DOS apps (think DOOM) use chained modes or bank flipping with their extensive use of VGA registers rather than direct framebuffer writes. Some chips are just more optimised at this than others and that's why the scores are not identical despite using the same bus and equivalent RAM chips.

The question is why are you comparing these two cards?
If you have both, try them and then decide what works best for you.
If you're looking for purchase advice then you should consider other cards aswell. 😀

The Serpent Rider wrote on 2020-01-29, 19:14:

Vlask setup isn't the fastest one, just convenient for ISA cards. […]
Show full quote

So I guess if 30% is the definition of huge..

Vlask setup isn't the fastest one, just convenient for ISA cards.

Even Vlask's benchmarks show a 30% difference between Mill2 PCI and S3Trio64 PCI (92 fps vs 71 fps).

67 vs 92 actually. And your experience might get even worse with a "noname" S3 Trio64V+.

Still, Vlask's charts can be misleading because the fast platform exaggerates the differences between cards as they'd perform on any normal DOS gaming system. I see lots of weirdness comparing his results to mine. For example, on my system:
-TNT2 M64 outperforms Voodoo3 2000 PCI
-ARK1000/2000 is only slightly slower than TNT2 M64 and faster than V3 2000.

His charts definitely have a place. It's neat to see how the chipsets perform with no restrictions and it's interesting to see how much headroom different chipsets have. But for the OP and most people building a DOS gaming system, they can get the wrong impression from looking at his results and expecting similar differences between graphic chipsets on their period-correct system.