All good Win-accellerators are to weak in DOS and vice-versa. So the best solution is 2Mb Cirrus, S3 or Trident card.

S3 Trio64+. I have one but they are nigh impossible to find.

You could also look for an ATI Mach32 (2MB max) or Mach64 (4MB max) but the Mach64 is generally insanely expensive.

The other option is to change to a motherboard that has PCI slots.

Babasha wrote on 2022-03-31, 19:29:

All good Win-accellerators are to weak in DOS and vice-versa. So the best solution is 2Mb Cirrus, S3 or Trident card.

My experince is that most S3 cards are pretty good in DOS and between ok and really good in windows.

The Cirrus Logic card I have now is limited to 1024x768 and maybe 60-70hz even with 2mb vram so its not that useful.

cyclone3d wrote on 2022-03-31, 19:40:

S3 Trio64+. I have one but they are nigh impossible to find.

You could also look for an ATI Mach32 (2MB max) or Mach64 (4MB max) but the Mach64 is generally insanely expensive.

The other option is to change to a motherboard that has PCI slots.

I have seen some threads about S3 Trio64 and Virge for VLB... and they are as common as unicorns.

Didn't think about the ATI-cards, should ad them to my watch-list.

Changing to PCI would be way to easy and ruin the whole idea of this system 😁

Vogon user MPE has tested a lot of VLB cards in both DOS and Windows.
Some of cards listed there do support 4MB (Diamond made S3 Trio64 for example), but they are rare.
Most VLB cards had 1MB - which was enough for most users (at the time), some had 2MB, and 4MB versions were late models and very, very rare.
4 MB cards became common much later, in PCI era, when VLB cards was long gone.

https://dependency-injection.com/vlb-vga-group-test/

phantom_pl wrote on 2022-03-31, 20:01:

Vogon user MPE has tested a lot of VLB cards in both DOS and Windows. Some of cards listed there do support 4MB (Diamond made S3 […]
Show full quote

Vogon user MPE has tested a lot of VLB cards in both DOS and Windows.
Some of cards listed there do support 4MB (Diamond made S3 Trio64 for example), but they are rare.
Most VLB cards had 1MB - which was enough for most users (at the time), some had 2MB, and 4MB versions were late models and very, very rare.
4 MB cards became common much later, in PCI era, when VLB cards was long gone.

https://dependency-injection.com/vlb-vga-group-test/

That test/review had lots of good information for me, thanks!

I know this is like chasing unicorns and that the 2mb cards are rare and the 4mb cards are more or less unobtainable 😀
But thats a part of the fun with these old systems.

Look for S3 864 or better. Price wise it's not too expensive but very good DOS and Windows speed. I'd say it's good compromise between price and performance.

I am using the Diamond Stealth 64 DRAM VLB in my 486 system with upgraded memory, runs without issues and has a pretty good output.

I have tested some VLB cards in system with a Am5x86 @160MHz for DOS apps.

In general the following accelerators can get a recommondation in DOS and Windows:

S3 864 / S3 868 / S3 968 / S3 Trio64 / S3 Trio64 V+ (all with 2MB DRAM or 2/4 MB VRAM)
ATI Mach32 2Mb VRAM / ATI Mach 64 2/4 MB VRAM
Trident TGUI 9440 2MB DRAM
ALG 2228 2MB DRAM
Cirrus Logic GD5434 2MB DRAM

In any case (for the cards above) make sure, you have 2MB RAM for 64Bit memory access or Memory interleave (Mach32) - only important for windows. For windows use make sure you have a (>)135MHz Hi- or Truecolor DAC. With Trio64 and TGUI9440 i observed spaghettized pictures on TFT screens. The best affordable, fastest card would be one with S3 864 like @kixs recommends. The S3 cards with DRAM are faster in DOS than the VRAM variants. In Windows the S3 VRAM cards are faster. With ATI cards the VRAM cards are equal to DRAM cards in DOS and faster in Windows.

The S3 Vision cards: 864/868 (DRAM) and 964/968 (VRAM) and Trio64 are 64Bit.
The older S3 cards 805 (DRAM) and 928(VRAM) and Trio32 are 32Bit, while the 805i has memory interleave with 2MB DRAM.
The S3 868/968 are nearly identical to the 864/964 with some additional video acceleration and zooming features.
The Trio64 hat the 864 core with integrated 135MHz RAM DAC.

For 1600x1200 @8bit with 60Hz or higher you need a 4MB card with 175MHz (or higher) RAM DAC. So you should look for: ATI Mach64 +Memory Upgrade, Diamond Stealth 64 (Video) VRAM +Memory Upgrade (was available with S3 964 and 135MHz DAC and as „Video“ with S3 968 and 220MHz DAC), ELSA Winner 2000 4MB version (was available with S3 928 and 964).
All of these cards are nearly unobtainalble. The best chance for you is to ask the user Madao, he developed a own S3 968 4MB VRAM card.

The chart is informative, but also a bit misleading.
The synthetic tests that kind of dont matter dominate the bars and skew the perceived qualities of the cards, while the 2 tests that actually matter - doom and quake 1 - have only minor contribution to the the overal bar size.
Separating the two groups can make it more clear.

This chart is DOS specific? No Win/GUI test ?

Windows gui tests are owned by the 64-bit s3 cards.
Keep in mind that there even 2x difference in performance is imperceivable.
MPE's blog has win gui tests. I can point you to some more i baked in another thread too.

In dos, if you focus on doom and quake 1 numbers the difference is within 2-4 frames in doom and under a frame in quake 1, for the upper 5 or so graphics adapters.

If you dont chase peak perf for bragging rights any of these 5 or so cards will be juuust fine.

In my experience, these DOS tests are quite informative for gaming applications. Especially the PC Player Bench reflects the reality quite well. E.g. the scrolling speed in SC2000 and other SVGA titles. For this reason the PCP VGA test will be similarly close to reality. The Chris3D Bench correlates similarly with the PCP Bench. The Quake test is very CPU limited on 486 VLB systems and not very meaningful for 486. The Doom engine is actually the only "real" application, although this scales with memory clock rather than other features and Bus width - Miro 20SV (S3 964) vs. Miro 20SD (S3 864) vs. Spea Mercury (S3 928) vs. Miro 8S (S3 805)

Attached is a chart normalized to ET4000AX ISA.

You realize you could simply use a P90 system with PCI graphics? 😜
(I understand that many people want to max out a certain platform but I never understood its appeal in retro.)

There's actually a pretty obvious source for superficial info on S3 chips:
https://en.wikipedia.org/wiki/S3_Graphics#Gra … ics_controllers
- Any card from the "vision" range will be pretty good already and maybe somewhat obtainable.
Same for Cirrus Logic which probably is the other more abundant supplier of that time:
https://en.wikipedia.org/wiki/Cirrus_Logic#Desktop
Beyond that, click your way through "VGA legacy" and the links to contemporary tests there.

High windows resolutions and color depths:
If you do have too much money, well, it happens. Then go buy a rare 4 MB VLB card.
Most will be 1 or 2 or 1 with the sockets or solder pads for 1 more. The usual screen back then was 14" and 640 x 480 or maybe 800 x 600, for mortals.

1 byte is 8 bit which would define 2⁸ = 256 different colors or shades of grey and video memory is (number of horizontal) pixels x lines x number of bits
Divide that bit count by 8388608 and you'll have it in MByte.
Screen Frequency is RAMDAC frequency divided by pixel count, e.g. IIRC, the Matrox Millennium had 220 MHz, divided by 1600 and 1200 that made 115 Hz. The original Mystique 'only' had 170 MHz, hence the later "220" edition.

Have fun!

@PC-Engineer

For synthetic tests to be meaningful one needs to build a mental map about what their results mean.
For example - if Superscape gives me 40 fps this is probably not enough for Doom class games, or if PCP Bench gives me 40 fps this may be enough for Quake 1 class games. But i better run the Doom and Quake 1 tests to verify that.
And that's why i meant with my early comment that Superscape, Chris, PCP, etc are kind of whatever.
But also - over time we figure-out the ration between synthetic and game tests and most people are able to follow.

About PCP Bench:
Similar to Quake 1 it is very FPU dependent.
This can be verified easily on 386 machines with co-proc on and off - delta is large.
That's why (as you also pointed out) Doom is better fit for 486 benchmarking, while PCP and Q1 are better fit for Pentium hardware.

---

@Cuttoon

Following your logic we should just go for DOSBox or other emulators and skip the physical side of things entirely.
I think PC-Engineer wrote somewhere "journey is the mission" and that's the spirit of the community (and our pointless activities).
; )

pshipkov wrote on 2022-04-03, 20:45:

Following your logic we should just go for DOSBox or other emulators and skip the physical side of things entirely.
I think PC-Engineer wrote somewhere "journey is the mission" and that's the spirit of the community (and our pointless activities).
; )

Not quite what I meant.
While I'm very much on board about older hardware over emulators, I don't quite get why roughly 9 out of 10 discussions are about overclocking, benchmarks and maximum specs.
To me, retro computing is about using a 486 because it's not an I7.
But why would I want it to be a Pentium?

I see. Ok.
I think it is a natural progression.
Most start on a path of just remembering and exploring the old stuff.
Over time the initial thirst is satisfied and we get more picky, which directly leads to building idealized rigs, which often leads to overclocking.
There are also arbitrary self-imposed limitations at play related to different hardware platforms. Some are into Pentiums, others into 386 and so on. To each their own.
Such forum members collected and profiled more "factual" information so they have more to share and tend to be more active in general, thus the louder chatter.
Or something like that.

pshipkov wrote on 2022-04-03, 07:13:

The synthetic tests that kind of dont matter dominate the bars and skew the perceived qualities of the cards, while the 2 tests that actually matter - doom and quake 1 - have only minor contribution to the the overal bar size.

Wouldnt call Quake important for VLB cards.

@pshipkov

I can understand your arguments very well and you are basically right that synthetic benchmarks are only indications and have to be put into perspective with experience.

That's why I chose exactly these benchmarks from Phil's suite. In my experience, they are (in sum) a good indicator for the real world performance. And they are accessible to everyone and thus reproducible. Speedsys, for example, is a very poor indicator.

I once did a CPU comparison on a Socket 3. The qualitative performance comparison of these benchmarks is almost identical to Windstone 95, which is a very realistic performance indicator for entire system with RAM and HDD speed.

By the way, I noticed that the PCP benchmark only scales very slightly with the FPU capabilities (compare SX/33 vs. DX/33 or iDX4 100 vs. POD), whereas Quake scales very strongly with FPU capabilities (486 vs. Pentium). But all benchmarks scaling very well with CPU clock. Sadly i did have no reference in my graphics card comparision, like Winstone95 for CPU comparison. With that fact in mind and that Quake is too hard limited by the 486 FPU i think my DOS graphics card comparison for 486 VLB graphic cards is a useful indication for real world DOS performance of these graphic cards.

@cuttoon

You are right, for 1600x1200x8bpp 2MB RAM are enough. But for the requirement of >60Hz at this resolution you need a 175MHz or higher clocked RAMDAC. The 135MHz types are not capable for this. The only cards with this class of DAC (175MHz or higher) are some of the 4MB capable cards, written above.

Although not top notch the Western Digital 1mb+ vlb cards always did respectably in both Windows and dos