That really depends on how and what you measure. DOS/other "dumb" ways of output and GUI acceleration are separate things.
Also you have to take into account that not every card is clocked at the same speed, which creates discrepancies like humble vanilla S3 Trio64 beating newer and more powerful (on paper) cards that clocked rather anemically.

My guess would be ET6100, followed by ET6000 - both are 128-bit.
All the others are 64-bit, aren't they?

Well no. Tseng ET6000/ET6100 has 32-bit bus - https://www.xfree86.org/3.3.6/tseng13.html

And once again, one of my Trio 64 is poking fun at so called 128-bit Tseng card in most of my tests, simply by virtue of being factory overclocked.

XGI Volari Z7/Z9/Z11 ? I don't know if they're 2D only.

ASPEED Technology AST2510 ?; https://www.vgamuseum.info/index.php/cpu/item … -aspeed-ast2510

The Serpent Rider wrote on 2025-06-23, 12:35:

Well no. Tseng ET6000/ET6100 has 32-bit bus - https://www.xfree86.org/3.3.6/tseng13.html

And once again, one of my Trio 64 is poking fun at so called 128-bit Tseng card in most of my tests, simply by virtue of being factory overclocked.

i would say by datasheet rated clock.
968: 50mhz cl1 vram with edo mode
trio64: 60mhz edo
trio64v2: 60mhz cl1 edo
cirrus5480: 100mhz sgram

noshutdown wrote on 2025-07-30, 08:41:

[...] […]
Show full quote

[...]

i would say by datasheet rated clock.
968: 50mhz cl1 vram with edo mode
trio64: 60mhz edo
trio64v2: 60mhz cl1 edo
cirrus5480: 100mhz sgram

That's memory clock, not VGA core clock. In any event, you can only 1:1 compare clock speeds within the same architecture. As soon as you look at different architectures (VRAM vs EDO DRAM vs SGRAM, as well as different core arch) scaling will be different. The highest number of MHz of any given clock does not necessarily mean a card will be fastest in a particular application.

But back to the question you haven't answered yet: what do you mean by "2d-only"?

Are we talking DOS? Or a particular version of Windows? Also, what resolution do you want to compare at? And in what benchmark?

If you reformulate your question as: "Which VGA card without 3D acceleration gives the highest score in benchmark A or most FPS in game B at resolution XY and settings Z?" we can give you an unambiguous answer (or at least discuss based on the same premises).

dionb wrote on 2025-07-31, 09:10:

Are we talking DOS? Or a particular version of Windows? Also, what resolution do you want to compare at? And in what benchmark?

If you reformulate your question as: "Which VGA card without 3D acceleration gives the highest score in benchmark A or most FPS in game B at resolution XY and settings Z?" we can give you an unambiguous answer (or at least discuss based on the same premises).

i would say to run a set of widely recognized tests to yield an overall rating, which shall include:
dos games and tests: both vga and svga, svga are expected to be vesa mode 101h(640*480*8bpp) though. x-vesa shall be included as it allows testing high resolution vesa modes in dos.
windows acceleration: winbench, wintune, winmark... and whatever you recognize. just remember to count dib dlit and directdraw blit in as they are more related to raw throughput and gaming performance.
windows games: i am more into 3d cards so i don't know about 2d games well, maybe starcraft? also software rendering 3d games at high resolution seems quite cpu bound and most stays at 256 colors, although counterstrike seems quite colorful and is probably in 16bpp mode.
also, windows98 is the recommended version as the cards drivers should be quite mature at that time, some cards seems to have no drivers for windows2000 and later.

as for resolution, i would pick three svga modes in addition to standard vga, to figure out a combined performance:
low: 640*480*8bpp
medium: 800*600*16bpp
high: 1024*768*24bpp(so the cards must have 4mb or more to enter this contest)

By my benchmark suit with tested mainly ISA and VLB card, I also thrown in some PCI ones for comparison. Best overall is the Tseng ET6000.

The Serpent Rider wrote on 2025-06-23, 12:35:

And once again, one of my Trio 64 is poking fun at so called 128-bit Tseng card in most of my tests, simply by virtue of being factory overclocked.

Can you provide more info about those tests?

As visible in the previous post, the difference between Trio64 and ET6000 is pretty high - hard to believe it can be compensated by overclocking.

noshutdown wrote on 2025-08-01, 03:42:

[....] […]
Show full quote

[....]

i would say to run a set of widely recognized tests to yield an overall rating, which shall include:
dos games and tests: both vga and svga, svga are expected to be vesa mode 101h(640*480*8bpp) though. x-vesa shall be included as it allows testing high resolution vesa modes in dos.
windows acceleration: winbench, wintune, winmark... and whatever you recognize. just remember to count dib dlit and directdraw blit in as they are more related to raw throughput and gaming performance.
windows games: i am more into 3d cards so i don't know about 2d games well, maybe starcraft? also software rendering 3d games at high resolution seems quite cpu bound and most stays at 256 colors, although counterstrike seems quite colorful and is probably in 16bpp mode.
also, windows98 is the recommended version as the cards drivers should be quite mature at that time, some cards seems to have no drivers for windows2000 and later.

as for resolution, i would pick three svga modes in addition to standard vga, to figure out a combined performance:
low: 640*480*8bpp
medium: 800*600*16bpp
high: 1024*768*24bpp(so the cards must have 4mb or more to enter this contest)

Three points:
1) you are describing three completely different use cases here: unaccelerated framebuffer performance (in DOS), 2D accelerated performance (in Windows) and software rendering of 3D (in Windows). It's to be expected that the results of the first two will be radically different, as frequently the cards with the best framebuffer performance have poor or absent acceleration, and as you say the latter is more a test of CPU than video card. If you want a singular answer of "the" fastest card, choose one.
2) you original question was about "2d -only" but you're mentioning Counterstrike and software 3D rendering. How does that tell you anything about 2D-only performance?
3) purely practically: you're describing at least ten benchmarks (and that's just taking a single game whenever you say "games") in four resolutions = 40 total benchmarks per card. Who exactly are you expecting to do all this work? Also here again - with that many different ways of testing you are not going to get a singular answer to your original question. If you want a clear answer, I strongly suggest choosing a single use case that actually matters to you.

Grzyb wrote on 2025-08-01, 10:54:

Can you provide more info about those tests?

PCI Video cards benchmark: Electric Boogaloo Edition

dionb wrote on 2025-08-01, 11:12:

1) you are describing three completely different use cases here: unaccelerated framebuffer performance (in DOS), 2D accelerated performance (in Windows) and software rendering of 3D (in Windows). It's to be expected that the results of the first two will be radically different, as frequently the cards with the best framebuffer performance have poor or absent acceleration, and as you say the latter is more a test of CPU than video card. If you want a singular answer of "the" fastest card, choose one.

true, what i mean by "fastest" is "best allrounder" that can meet the requests of all cases well.

2) you original question was about "2d -only" but you're mentioning Counterstrike and software 3D rendering. How does that tell you anything about 2D-only performance?

i do mean cards without any features for 3d, but don't "3d games in software rendering mode" like quake, unreal and counterstrike have cpu doing all 3d rendering work and then simply deliver the rendered image to the video card for output? so why do you think that they make use of any 3d features from that video card?

3) purely practically: you're describing at least ten benchmarks (and that's just taking a single game whenever you say "games") in four resolutions = 40 total benchmarks per card. Who exactly are you expecting to do all this work? Also here again - with that many different ways of testing you are not going to get a singular answer to your original question. If you want a clear answer, I strongly suggest choosing a single use case that actually matters to you.

in fact there are only a few cards left for candidates of fastest 3d-free card, while others are already known to be mediocre at best.
and the tests can be chosen to save some time while remaining recognized. also, some games have fixed resolutions or color depths so they don't have to be run for many times.

The Serpent Rider wrote on 2025-08-01, 11:31:

PCI Video cards benchmark: Electric Boogaloo Edition

thanks for the tests, while they don't involve much true color performance or windows acceleration(except that c&c2 which doesn't have precise fps number), they surely help eliminating some contestants that already seem incompetitive.
and just why is the cirrus5430 this slow? i know it has 32bit ram width but so is the trio32, and its almost 3 times as fast as the 5430.

and just why is the cirrus5430 this slow? i know it has 32bit ram width but so is the trio32, and its almost 3 times as fast as the 5430.

Apparently 54M30 specifically and also 5440 are just that cursed for whatever reason. Vlask did some benchmarks on vgamuseum and normal 5430 is faster, although still slow in general.

The Serpent Rider wrote on 2025-08-04, 23:28:

and just why is the cirrus5430 this slow? i know it has 32bit ram width but so is the trio32, and its almost 3 times as fast as the 5430.

Apparently 54M30 specifically and also 5440 are just that cursed for whatever reason. Vlask did some benchmarks on vgamuseum and normal 5430 is faster, although still slow in general.

i checked it out and while there is no datasheet for 54m30, the difference between it and 5430 are supposed to be:
the 5430 is in 208pin package just like the more expensive 5434 and 5436, and has 64bit ram width when equipped with 2mb. its cut down from 5434 in 32bit bitblt, 86m pixel rate , and 2mb of maximum ram. still its significantly slower than 5434 and only a bit faster than 5429 though.
the 54m30 and 54m40 are in 160pin package similar to the old 542x series, and have both 32bit ram width and bitblt. not sure if they are just a slightly modified 5429 to support pci bus.

Yes, it's very likely that those chips are just 5429 with some minor tweaks to DAC (no interlaced mode for 1024x768) and PCI bus, but with no performance to back that up.

A lot of Diamond Stealth 64 S3 968 come with 60NS VRAM but the bios is programmed for the 70NS models.
If you have 60NS chips you can modify the BIOS (search Unsenet 1990's era posts) to increase the memory Mhz to >60+mhz. I remember after doing this it was faster than my TRIO64 under DOS benchmarks, faster in VESA 640x480+ modes as well. For raw bandwidth I remember the Dos benchmark VSpeed went from around 89MB/S to 100MB/S.
It was probably true for #9 and Spea S3-968 models as well. The cards (depending on chip availability) either came with 70 or 60ns chips but the bioses were only made to support 70ns ram.

Riikcakirds wrote on 2025-08-20, 23:32:

A lot of Diamond Stealth 64 S3 968 come with 60NS VRAM but the bios is programmed for the 70NS models.
If you have 60NS chips you can modify the BIOS (search Unsenet 1990's era posts) to increase the memory Mhz to >60+mhz. I remember after doing this it was faster than my TRIO64 under DOS benchmarks, faster in VESA 640x480+ modes as well. For raw bandwidth I remember the Dos benchmark VSpeed went from around 89MB/S to 100MB/S.
It was probably true for #9 and Spea S3-968 models as well. The cards (depending on chip availability) either came with 70 or 60ns chips but the bioses were only made to support 70ns ram.

that sounds interesting, but i would rather stay with datasheet specs to be safe, which says that 964 supports normal edo mode up to 60mhz, while 968 also supports single cycle mode up to 50mhz, which is supposed to be faster.
also note that while the trio64v2 is a bit faster in most dos games(which don't involve truecolor and high resolutions), its datasheet says it would only support up to 1024*768 in highcolor and 800*600 in truecolor, which seems too low to compete for the fastest 2d card. guess its simply not fast enough to support higher resolutions for not using vram.
still the 968 seems like a pinnacle of 2d cards to me, only matched by the matrox millenium which has some useless primitive 3d features and is therefore not a pure 2d card.

S3 Trio 64V2/DX is faster in high color than S3 968 too. External DACs on 968 are better, but it doesn't matter much if you're stuck with only 2 Mb, which is often the case for S3 Vision 968 cards.