VOGONS


PCI Graphics Roundup - DOS

Topic actions

First post, by clueless1

User metadata
Rank l33t
Rank
l33t
PCI graphics cards 2.jpg
Filename
PCI graphics cards 2.jpg
File size
2.59 MiB
Views
13219 views
File comment
The Collection
File license
Fair use/fair dealing exception

This thread is dedicated to benchmarking PCI graphics cards in DOS games. As I test new cards, I will add them to the results.

The testbed is a Packard Bell socket 5 with Triton chipset and 32MB 72-pin EDO. No L2 cache on the motherboard. Integrated CL-GD5430 with 2MB. I will be focusing on VGA performance at 100Mhz, as at this speed most SVGA games don't run very smoothly. At 200Mhz I will be focusing on SVGA results, since games at SVGA resolutions for the most part are very playable on a 200MMX. Also, I will test each card in real DOS on a P3-933 to see how the cards perform with less of a CPU bottleneck. Hint: some cards scale well up to 933Mhz, other cards that perform well at 100 and 200Mhz don't scale as well up to 933, so you see different rankings at different speeds.

I'm reporting raw FPS results as well as performance percentages. So in a particular benchmark, the highest scoring card gets 100% and each of the other cards gets scored on their percent score compared to the top result. This way, we can average the results in the end without having to worry about the weighting of the numbers being inconsistent between the different benchmarks.

Games used:
Wolfenstein 3D (VGA)
Doom (VGA)
Quake (VGA: 320x200, 360x480 / SVGA: 640x480)
Duke Nukem 3D (VGA: 320x200 / SVGA: 640x480, 800x600)
Descent II (VGA: 320x200 w/cockpit view, 320x400 no cockpit / SVGA: 640x480 w/cockpit view, 640x400 no cockpit)

Pentium 100 VGA Results:

P100VGA.png
Filename
P100VGA.png
File size
80.25 KiB
Views
12955 views
File license
Fair use/fair dealing exception

Let's see what happens when we add 100Mhz and MMX to the mix and crank up the resolutions.

Pentium Overdrive MMX 200Mhz SVGA Results:

P200SVGA.png
Filename
P200SVGA.png
File size
69.17 KiB
Views
12954 views
File license
Fair use/fair dealing exception

VGA Performance at 200Mhz:

P200VGA.png
Filename
P200VGA.png
File size
81.46 KiB
Views
12954 views
File license
Fair use/fair dealing exception

And lastly, how the cards did in SVGA games at 100Mhz:

P100SVGA.png
Filename
P100SVGA.png
File size
66.3 KiB
Views
12954 views
File license
Fair use/fair dealing exception
Last edited by clueless1 on 2017-02-10, 16:32. Edited 5 times in total.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 1 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

To reduce any CPU bottlenecking and see how these cards perform with more processing power, I tested all these cards in a Pentium III 933Mhz system. The system is a Dell Dimension 4100.

First, let's look at VGA games:

P933VGA.png
Filename
P933VGA.png
File size
76.79 KiB
Views
12956 views
File license
Fair use/fair dealing exception

Now for SVGA games:

P933SVGA.png
Filename
P933SVGA.png
File size
63.98 KiB
Views
12956 views
File license
Fair use/fair dealing exception

VGA vs. SVGA
I compared overall performance of each card at VGA and SVGA. I also noted in red the cards that had a significant drop (3% or higher STDEV) in SVGA performance compared to its VGA:

VGAvsSVGA.png
Filename
VGAvsSVGA.png
File size
70.59 KiB
Views
12955 views
File license
Fair use/fair dealing exception

The most consistent-performing cards across all three CPU speeds are the TNT2 M64, ARK2000PV, and CL-GD5446.

Reviews and Overall Ratings -

TNT2 M64 (1999)
Highs:
-native VBE3.0 means no need for UniVBE to run Quake in SVGA.
-at 100Mhz and 200Mhz, VGA performance is within 0.2% of the lead.
-at 933Mhz its VGA performance is the best.
-at 100Mhz, 200Mhz, and 933Mhz, its SVGA performance is the best.
-very good compatibility with DOS games.
-relatively inexpensive and easy to find.
Lows:
-being made in 1999, it's not period-correct for a DOS PC, if that's important to you.

ARK2000PV (1995)
Highs:
-overall VGA speed at 100Mhz and 200Mhz is best by a negligible amount.
-SVGA performance at 100Mhz and 200Mhz is within 2.5% of the lead.
-at 933Mhz, SVGA and overall performance is 2nd only to TNT2 M64.
-of the cards tested, it's the fastest period-correct card for an MS-DOS gaming machine.
-very good compatibility with DOS games.
Lows:
-natively supports VBE1.2, so requires a TSR (UniVBE, 13-17KB TSR) for Quake SVGA.
-hard to find and expensive unless you get lucky.

Voodoo3 2000 (1999)
Highs:
-native VBE3.0 means no need for UniVBE to run Quake in SVGA.
-VGA and SVGA performance within 5% of the leaders at 100Mhz and 200Mhz.
-2nd fastest VGA performance and 3rd fastest SVGA performance at 933Mhz.
-good compatibility with DOS games (see lows).
Lows:
-some issues with LCD monitors according to https://gona.mactar.hu/DOS_TESTS/
-Wing Commander 3 requires an alternate video mode setup for movies to work.
-SVGA performance gets gradually worse as CPU speed increases.
-being made in 1999, it's not period-correct for a DOS PC, if that's important to you.

Matrox Millennium II (1997)
Highs:
-VGA and SVGA performance within 5% of the leaders at 100Mhz and 200Mhz.
-2nd fastest period-correct card of the group.
-native VBE2.0 means no need for UniVBE to run Quake in SVGA.
Lows:
-numerous compatibility issues according to https://gona.mactar.hu/DOS_TESTS/
-I noticed Descent2 800x600 has a scrambled display, and Chris's 3D Benchmark in SVGA crashes the system.
-limited overhead as CPU speed increases, especially in SVGA.

ATI Rage 128 GL (1998)
Highs:
-Very fast VGA performance, within 5% of the leaders no matter what CPU speed.
-native VBE2.0 means no need for UniVBE to run Quake in SVGA.
Lows:
-numerous compatibility issues according to https://gona.mactar.hu/DOS_TESTS/
-SVGA performance lags behind, becoming much worse as CPU speed increases.
-made in 1998, it's not period-correct for a DOS PC.

S3 Virge/GX (1996)
Highs:
-fastest of the S3 cards I tested.
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-VGA performance within 4% of the leaders at 100Mhz and 200Mhz.
-though it requires a TSR for VBE2.0, its TSR (S3VBE20.EXE) uses only 6KB of memory compared to 13-17KB for UniVBE.
-very good performing period-correct option.
-inexpensive (usually less than $15) and easy to find.
Lows:
-natively supports VBE1.2, so requires a TSR for Quake SVGA.
-some S3 cards have high brightness issues that cause blacks to appear grey.
-limited performance overhead as CPU speed increases, causing it to fall further behind competition on faster systems.

S3 Virge/DX (1996)
Highs:
-tied with V2/DX as 2nd fastest S3 card.
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-VGA performance within 5% of the leaders at 100Mhz and 200Mhz.
-Diamond includes built-in VBE2.0, so no TSRs are needed to run Quake SVGA.
-decent performing period-correct option.
-inexpensive (usually less than $15) and easy to find.
Lows:
-some S3 cards have high brightness issues that cause blacks to appear grey.
-even more limited performance overhead as CPU speed increases, compared to the Virge/GX.

S3 Trio64V2/DX (1996)
Highs:
-tied for 2nd fastest S3 card in theses tests.
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-VGA performance within 5% of the leaders at 100Mhz and 200Mhz.
-though it requires a TSR for VBE2.0, its TSR (S3VBE20.EXE) uses only 6KB of memory compared to 13-17KB for UniVBE.
-decent performing period-correct option.
-inexpensive (usually less than $15) and easy to find.
Lows:
-natively supports VBE1.2, so requires a TSR for Quake SVGA.
-some S3 cards have high brightness issues that cause blacks to appear grey.
-even more limited performance overhead as CPU speed increases, compared to the Virge/GX.

S3 Virge/325 (1996)
Highs:
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-Overall performance at 100Mhz within 4% of the leaders.
-VGA performance within 6% of the leaders at 100Mhz and 200Mhz.
-though it requires a TSR for VBE2.0, its TSR (S3VBE20.EXE) uses only 6KB of memory compared to 13-17KB for UniVBE.
-period-correct option.
-inexpensive (usually less than $15) and easy to find.
Lows:
-natively supports VBE1.2, so requires a TSR for Quake SVGA.
-some S3 cards have high brightness issues that cause blacks to appear grey.
-even more limited performance overhead as CPU speed increases, compared to the Trio64V2/DX.
-over 20% behind the leaders overall at 933Mhz (very limited overhead as CPU speed increases).

S3 Trio64V+ (1995)
Highs:
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-Overall performance at 100Mhz within 6% of the leaders.
-VGA performance within 8% of the leaders at 100Mhz and 200Mhz.
-though it requires a TSR for VBE2.0, its TSR (S3VBE20.EXE) uses only 6KB of memory compared to 13-17KB for UniVBE.
-period-correct option.
-inexpensive (usually less than $15) and easy to find.
Lows:
-natively supports VBE1.2, so requires a TSR for Quake SVGA.
-some S3 cards have high brightness issues that cause blacks to appear grey.
-very limited performance overhead as CPU speed increases (over 12% behind leaders overall at 200Mhz).
-almost 30% behind the leaders overall at 933Mhz (very limited overhead as CPU speed increases).
-the slowest S3 card in these tests.

Cirrus Logic CL-GD5446 (1996)
Highs:
-within 3% of the leaders at 100Mhz.
-within 6% of the leaders at 200Mhz.
-SVGA performance does not drop proportionately more compared to VGA like many of the other cards.
-excellent game compatibility according to https://gona.mactar.hu/DOS_TESTS/
-combination of good performance and period-correctness.
Lows:
-natively supports VBE1.2, so requires a TSR for Quake SVGA.
-must use SDD53a or UniVBE for VBE2.0, which requires more memory (13-17KB) than S3VBE20.
-limited performance overhead as CPU speed increases (16% behind the leaders at 933Mhz vs only 6% at 200Mhz).
-a little pricey (about $25) unless you get lucky.

CL-GD5430(integrated chip, not rated)

My personal recommendations -
If period-correctness is not important to you, the TNT2 M64 is the obvious choice. In VGA games, it's never more than 2.5% slower than any other card, no matter what CPU speed. In SVGA games, it's always the fastest, which makes it a great choice for a fast DOS PC that will be playing demanding games in SVGA like System Shock, Screamer 1/2, Descent II, USNF, etc. Something like a P233MMX or K6-2.

If you want a period-correct graphics card for your DOS gaming machine, my top-3 are:
1) ARK2000PV - fastest period-correct card in VGA and SVGA, nice image quality, good compatibility. Hard-to-find and usually expensive.
2) S3 Virge/GX - very fast in both VGA and SVGA, and consistent performance among all the games tested. Famous compatibility and nice image quality. Easy-to-find and inexpensive.
3) CL-GD5446 - performs very similar to the Virge/GX except in Quake, where it is significantly slower in 360x480 and 640x480. Nice image quality and excellent compatibility. Easy-to-find, but a little more expensive than the Virge/GX.

For slower Pentiums that will only be playing games in VGA resolutions, ALL of the cards perform fine.

The Millennum II and Rage 128 GL have the worst compatibility ratings, and I've personally experienced lots of compatibility issues with the Millennium II. The Rage 128 GL also has very poor SVGA performance. Therefore these two cards are on my "not recommended" list.

Last edited by clueless1 on 2017-02-11, 18:20. Edited 9 times in total.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 3 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t
sprcorreia wrote:

Nice work. Excellent starting point for choosing a card.

Thanks. All I'm really missing to make this comparison complete are:

CL-GD5446
ET6000
ET4000

I think those are the only remaining chipsets that are considered top DOS performers. Unfortunately, they are all too pricey on ebay at the moment. Even the GD5446 is >$20. Just have to be patient until one comes up with a low price.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 4 of 46, by Neco

User metadata
Rank Member
Rank
Member

I'm kind of confused why these cards/systems are not hitting 60fps on DOOM in your first chart.
I remember that game running silky smooth on 486 machines. Even Duke3D results seem weird to me, I could play that game even in Windows and it was pretty smooth all around from what I remembered.

Reply 5 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t
Neco wrote:

I'm kind of confused why these cards/systems are not hitting 60fps on DOOM in your first chart.
I remember that game running silky smooth on 486 machines. Even Duke3D results seem weird to me, I could play that game even in Windows and it was pretty smooth all around from what I remembered.

Check Phil's VGA Database - over 700 results:
Phil's Ultimate VGA Benchmark Database Project
Doom 60FPS doesn't happen until Pentium 133 speeds.

Back when I had a 486 and did not know what a Pentium felt like, I thought Doom was silky smooth too! But typical framerates for a 486DX2/66 in Doom is 25 fps. 🙁

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 6 of 46, by jade_angel

User metadata
Rank Member
Rank
Member

Doom has an internal framerate limit of 35FPS, which a 486DX4/100 will often reach with a decent VLB or PCI graphics card, so on a high-end 486, Doom actually is silky smooth - but you can feel the difference between a DX2/66 and a 5x86/133 during big fights, for sure.

To clarify, Doom will benchmark out above 35FPS, but in actual play, anything faster will get clipped down to 35FPS, or such is the best of my understanding.

Main Box: Macbook Pro M2 Max
Alas, I'm down to emulation.

Reply 7 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t
jade_angel wrote:

Doom has an internal framerate limit of 35FPS, which a 486DX4/100 will often reach with a decent VLB or PCI graphics card, so on a high-end 486, Doom actually is silky smooth - but you can feel the difference between a DX2/66 and a 5x86/133 during big fights, for sure.

To clarify, Doom will benchmark out above 35FPS, but in actual play, anything faster will get clipped down to 35FPS, or such is the best of my understanding.

Right, that's the way I understand it too. Kind of like leaving vsync on on modern games, so framerate is capped at 60 instead of maxing out, then turning it off when you run a benchmark to get the maximum framerate the card can output.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 8 of 46, by lazibayer

User metadata
Rank Oldbie
Rank
Oldbie

The ARK2000PV is said to have 64bit memory bitwidth, but with 4x 8bit chips it only gets 32bit. I wonder if it would overthrow M64 in every test once it had the empty memory bank populated. It uses 256k x 8bit DIP chips by the way.

Reply 9 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t
lazibayer wrote:

The ARK2000PV is said to have 64bit memory bitwidth, but with 4x 8bit chips it only gets 32bit. I wonder if it would overthrow M64 in every test once it had the empty memory bank populated. It uses 256k x 8bit DIP chips by the way.

Yeah, I'm curious about that too. Thanks for the info on the chip. I'm having a hard time finding it, though. Looks like it would be 24-pin, just counting the lead holes in each slot, and I can't find anything with 24 leads. And it would have to be 45ns or faster to match the soldered mem. If you (or anyone else) can post a link to to buy these chips, I'd appreciate it.

edit: looks like these would be 256KB chips: 32kbit x 8bit, 24-pin, 45ns DIPs.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 11 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Read that thread before, but I went back through it and noticed this quote by Anonymous Coward:

According to information I have seen in the past, the performance of the VGA core in both the 1000PV (32-bit) and 2000PV (64-bit) is supposed to be identical. I think I read that the VGA core on the 2000 is only 32-bit. Upgrading the 1000PV to 2MB gives no speed advantage over 1MB as it does not use 32-bit memory interleaving, but 2MB does provide extra resolutions in Windows.

which suggests upgrading the RAM probably won't change performance. Just as well, as that memory is proving difficult to find at all, let alone for a good price.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 12 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Updated the results in the first post to include performance consistency (standard deviation).

Also added results in the 2nd post on a P3-933 to reduce CPU impact on the scores. There's definitely more separation between results, and the TNT2 M64 overtakes the ARK2000PV with more CPU to play with.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 13 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Got an ET6000 and CL-GD5446 on the way. If they are functional as advertised, I'll add them to the results in the first two posts. Made offers on both, and ended up spending a combined $30 shipped for the two. No photo on the CL, so who knows what condition it's in.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 14 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Added results for the CL-GD5446 to the first two posts. Performance is typically somewhere between the Virge GX and Trio64 V2/DX. The ET6000 was DOA, but I got a refund, so no harm, no foul. I'll keep looking for good deals on this card so I can eventually add it to the chart.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 15 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Added Reviews/comments on each card at the end of the 2nd post.

edit: As I collect more PCI cards, I will update the benchmark results. Hoping for an ET4000/W32p and ET6000 (the only remaining cards I'm willing to pay for), but if I get any freebies I'll test them too. Any other cards I should be considering?

Last edited by clueless1 on 2017-02-04, 18:08. Edited 1 time in total.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 16 of 46, by BSA Starfire

User metadata
Rank Oldbie
Rank
Oldbie

Very nice set of reviews. Thanks for the work 😀

286 20MHz,1MB RAM,Trident 8900B 1MB, Conner CFA-170A.SB 1350B
386SX 33MHz,ULSI 387,4MB Ram,OAK OTI077 1MB. Seagate ST1144A, MS WSS audio
Amstrad PC 9486i, DX/2 66, 16 MB RAM, Cirrus SVGA,Win 95,SB 16
Cyrix MII 333,128MB,SiS 6326 H0 rev,ESS 1869,Win ME

Reply 17 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t

Added Diamond Stealth 3D 2000 Pro results with Virge DX chipset to the first two posts. This is a nice card! It has VBE2.0 built in, which sets it apart from the rest of the S3 cards. It performs nearly identically to the Trio64V2/DX, a little slower than the Virge GX. So the GX is still the fastest S3 card in these results.

Also, since I added the reviews at the end of the 2nd post, I removed a lot of redundant commentary. 😀

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks

Reply 18 of 46, by Shadoron

User metadata
Rank Newbie
Rank
Newbie

I would love to see a Cirrus Logic Laguna CL-GD5464 go up against the rest of your PCI cards. I picked one up from eBay for 10 bucks a month ago. They are supposed to have Cirrus Logic's fastest 2d core ever made, before they got out of the video card market. Info on them is pretty rare, the only old review I found stated that the 3d section of the card was only "ok" (compared to what though?). It has VBE2.0 built in.

Reply 19 of 46, by clueless1

User metadata
Rank l33t
Rank
l33t
Shadoron wrote:

I would love to see a Cirrus Logic Laguna CL-GD5464 go up against the rest of your PCI cards. I picked one up from eBay for 10 bucks a month ago. They are supposed to have Cirrus Logic's fastest 2d core ever made, before they got out of the video card market. Info on them is pretty rare, the only old review I found stated that the 3d section of the card was only "ok" (compared to what though?). It has VBE2.0 built in.

I'll keep my eye out for one, preferably free. 😉

I seem to remember forum member BSA Starfire having one of these and reporting disappointing DOS performance.

BTW, there's still a sliver of hope for an ET6000 review. I sent the DOA card I bought to a Vogons buddy to try and repair it. I gave it my best shot, but there's only so much you can do with an iron. He's got better equipment! So I'm hopeful.

The more I learn, the more I realize how much I don't know.
OPL3 FM vs. Roland MT-32 vs. General MIDI DOS Game Comparison
Let's benchmark our systems with cache disabled
DOS PCI Graphics Card Benchmarks