Thank you for all your work! I recently purchased a 486 MB to build a 486 retro computer. It has only ISA and VLB but supports dx4. I started panicking that it didn’t have PCI. Your benchmarks put my choice at ease. Thanks.
The trick with VLB vs PCI in today's market is the price of the cards themselves, with PCI examples of particularly sought-after mid-90's chipsets on VLB cards like the S3 968/Trio 64, ATI Mach 64, Tseng ET4000 and the Ark Logic parts pushing $200-$500 compared to $30-60 for PCI examples of cards with the same chipset/RAM configuration. These benchmarks also show that both synthetic and gaming benchmarks on socket 3 systems are almost all CPU/platform bound even with a higher-end 486 DX4/100 CPU in play. You also have the CPU cache method (WT vs faster WB), mainboard cache configuration, wait-state/memory timing, and any dividers in play related to VLB/PCI bus speeds depending on how you run your CPU. There are also the advanced features in the Cyrix 5x86 chips that can make them performance competitive with the Pentium OD chips in games like Quake that are highly FPU sensitive, where the Intel 486 class and AMD 586 chips fall down. The best bang for the buck in VL bus video cards would be something like a Cirrus Logic 5426/8/9 or a Trident 9440x VLB card for $50-60, however a $20 PCI based S3 Virge or Trio 64 in basically any flavor will exceed their performance in DOS and early Windows gaming performance. Having the PCI option gives you access to early 3D cards that will boost performance in games like Quake, including the Nvidia RIVA 128/TNT, 3dfx Voodoo 1/2, Rush and Banshee cards, all for less money than poorer performing "high-end" VLB cards. Compatibility with VBE/VESA modes for higher resolution DOS support is also a major factor to consider in selecting a card for gaming in your 486, even some of the high dollar VLB chipset cards mentioned previously have compatibility issues with some popular games which makes their relatively high performance somewhat pointless if it doesn't work well with your favorite game (gona.mactar.hu/DOS_TESTS/). ATI, Matrox and the Tseng ET cards can all be tricky in this regard in either VLB or PCI flavors.
Yes, I would agree with all of that. It depends if you are in it for collecting or using the parts of course. I honestly don't know whether I am in it for one or the other or both. I do have a big collection, but mostly I bought the parts very cheap or many years ago when the parts were not so expensive.
Thank you for posting this video, it was very interesting. After watching this I decided to run some benchmarks on my own system, as it has a (somewhat rare) Plug 'N Play 486 motherboard with PCI, from very late in the life of that CPU architecture - 1996. This may have a later and more advanced PCI implementation, as it supports PCI features such as CPU to PCI write cache, burst writes, etc. The system specifications are otherwise similar, including a DX4-100. The card I'm using is a PCI Diamond Stealth 2000 4mb (S3 Virge DX/GX). This achieves 64.2 fps in 3DBench, 43.7 fps in Chris, and 11.8 with the SVGA version. Similar results were had when I swapped that out for a PCI TNT 16mb card, so I think this might be the ceiling for the DX4-100, or at least my machine. Neither card showed any artifacting and completed all benchmarks (Player, Chris, 3DBench, etc). So in summary, I think you are spot on with your assessment of the early PCI implementation being (at least part of) the problem.
Your tests reflect what I remember from the past: As long as you had a VLB or PCI VGA card, you didn't have any issues with DOS gaming. None of us had a use for more than 1 or 2 MB of video RAM either, as only 14" monitors were kind of affordable. The real differences came with Video for Windows. I bought an ATI Mach64 VT in early '96 for my 486 specifically because it was one of the first cards to support hardware YUV video acceleration. With a Windows software decoder using that feature, I was able to watch MPEG1 videos in playcard size and was actually pretty happy with it. I still have those two CD-I discs that I bought '95 on an Amiga fair in Cologne and played a billion times on that card: "Star Trek II - Wrath of Khan" and "Total Recall" ;-)
Nice job! Very in depth! About the amount of video memory: Some graphics chipsets actually benefit from having extra video ram apart from the obvious higher color depths at higher resolutions. For instance: I have a rare (in vlb form) Cirrus Logic Alpine card (I think it's a CL-GD5434 but I never manage to memorize those Cirrus numbers ;-) ) that actually doubles the width of the internal memory interface when equiped with 2 megs of video ram instead of the standard 1 meg. So out of the box it features a 32-bit wide memory bandwidth but if you put an extra 1 meg in it, it doubles the width of the bus to 64-bit and that really gives it a nice speed boost! I expanded my card to 2 megs and this thing now really flies for a 1994 vlb graphics card on my P5 Pentium 60 (yes, it's a Pentium with VLB!)👌 It's not nearly as fast as an ET6000 though but that card is a lot newer of course.. 19:03 : You forgot to put on your 3D glasses! ;-) Graphics cards and their development over time keep fascinating me, especially during the era of the cards you reviewed here. Developments just exploded during the nineties. A very interesting subject indeed!
I do love having my biases confirmed. PCI on a 486 = no real advantage. Just stick with VLB for stability. (Kind of ironic given vlb was a bit of a bodgey stop-gap) My prejudices have been upheld! Thumbs up!
Yeah. This was the dream back then, but with the benefit of hindsight, crazy combo VLB/PCI boards and things like this were always VIA-based designs. And the thing I learned about VIA is that they were always quick to adopt the bleeding edge, and so those boards were perpetually a full of tray of half-baked cookies. There are so many boards I wanted (or still want), like Super Socket 7 stuff and all that.. all very tempting, but those features were thin on the ground for a reason.
Nice benchmarks. Just was hoping to see a cost reduced version Trio 32 included to see how much of a performance impact it has compared to the Trio 64.
Is this on an AMD DX4 100 100NV8T (8kb write-through cache) ? Can't get 3D bench beyond 51 on mine with an S3 Virge PCI card. (noticed yours goes to 62.5)
My Milleniums (MIL/2 and MIL2P/4, latest BIOS on both) also failed 640x480 Chris 3D test. ET-6000 sometimes was slightly faster than both Millenium and Virge.
I know this was a couple years ago, but maybe you still read recent comments. I think there is a limitation of you board in play. I have a Zeos Martin motherboard with 16bit ISA and VLB running a DX2. With a Diamond Stealth VLB I get 65.6 FPS. On my other machine I have a SuperMicro P5STE motherboard with 16bit ISA and PCI running a DX4. With a Matrox Millennium 2 I get 152.5 FPS. It isn't a particularly fast card, but it is twice the results of any of the PCI cards you tested and I'm sure many of those cards are faster. So that's what leads me to believe you motherboard might be the limiting factor. Just a thought. Thanks for all this great content.
Very interesting. I wonder what could be the reason for scores that high. I've tried quite a lot of boards by now and don't recall seeing anything like that, at least not without overclocking.
@@PCRetroTech I just ran them again. In the 640x480 PC Player I got 20.9. Not sure if that is decent. I ran the 3D again and got 153.3. In Quake I got 15.7 FPS. In Chris's 3D I got 147.9. and in Doom I got 936 realticks (don't think I got the right term). This machine isn't overclocked. I also tried a Pentium Overdrive 200 and got the same scores, within a frame or two.
@@robertmerrill9849 The highest score I have ever gotten with Quake on a DX4-100 without overclocking was about 11.7 fps. I've tried many boards. There's no way you can get the same score with a Pentium Overdrive because the Pentium had an insanely good floating point unit, which Quake really loves. So something very odd is going on there. I don't really have an explanation for the scores you are getting. If you can get 15.7 fps without overclocking on a stock DX4-100 then you can break the world record very easily with an overclocked AMD 5x86-133.
@@PCRetroTech My mistake for not being clearer. I have two setups and the one that got the 15.7 was a Pentium Overdrive with a PCI Matrox Millennium 2. I will have to take the other setup and replace the DX2 with the 4 and try it there.
VLB card speed depends on bus speed. At 40 or even 50 mhz VLB card can be faster than PCI. I overclocked amd 5x86 133 to 150 (50*3) and CL-GD5429 VLB was much faster.
Can you quantify the "was much faster" statement - what were your benchmark scores stock (4 x 33), at 160 mhz (4 x 40) and at 150 mhz (4 x 50) in Quake and Doom with that GD5429, presumably with the VL bus at a 1:1 ratio with the FSB? Curious how your results line up with the CL GD5428 VLB and CL GD54M30 PCI card results identified here, which nicely bracket the 5429 in performance / features. I've had little trouble running most PCI graphics cards at 40 mhz on the PCI bus when paired with an IBM/Cyrix 5x86c or AMD 5x86 CPU, usually to hit a 50 mhz FSB you are going to sacrifice raw clockspeed with a lower multiplier to remain stable unless you get a really nicely binned chip, so it's not just a straight gain of performance overall vs the 40 mhz FSB option.
@@TheYuppiejr The dx4 120 with 3x40 MHz was called the Pentium killer by us kids, but then came quake and all the 486 went into the bin... Never liked quake, because it destroyed my dream machine ;)
How are you having trouble finding Matrox drivers? They still have a website, and to this day, you can still download drivers for the Millennium and Mystique cards, going all the way back to the usual round of DOS applications and Windows 3.x.
The v7-Vega vlb is also a CL GD5428. Sad that vlb cards are hard to find nowadays. I have a trident 9440cxi vlb and I have to admit that the performance were not too bad, a bit under a gd5428 vlb.
The mystery card looks like a Datapath Horizon, you'll notice it has a universal 3.3v/5v pci connector which means it 'should' work in both types of slot but i've found these newer pci cards tend not to work in older 5v only pci systems. The Intel chip no the board is a pci bridge allowing two graphics cards to talk through the one pci bus connector.
@@PCRetroTech no probs, here is a link to some details www.vgamuseum.info/index.php/component/k2/item/717-datapath-horizon-2s VGA Museum is a good source for old card info.
Thanks for the benchmarks. Too bad it's been done on VIA chipset motherboard as those use VLB2PCI bridge chip (VT82C505) and as such, don't provide as much performance as true PCI chipsets (e.g. SIS 496/497). On top of possible instabilities as demonstrated in the video.
I'm at 17:13 - well, they don't have to support 87Hz. This 87Hz is interlaced (so it's effectively "half the resolution" for each frame, or "two frames" making up one), for monitors (and RAMDACs) that couldn't even do 60Hz at those resolutions. Almost any standard 14-15" CRT should do fine here. On a side note: my BenQ GW2470 can display those modes though it displays a perpetual "out of spec" message in the OSD. 26:56 - pop out of nowhere and suddenly gives better performance - This may be linear frame buffer. Do those cards support VBE 2.0+? I remember that Tseng ET6000's BIOS does.
Yeah, was going to say that same thing. 87Hz is actually more like 43Hz, but interlaced. It was pretty typical for CRTs of that era to support 640x480 @ 60Hz, maybe 75Hz; 800x600 @ 56Hz, maybe 62Hz; and 1024x768 interlaced @ 87Hz. When you saw "1024x768 NI" on a monitor spec sheet, that was something special.
@@PCRetroTech Alternatively, you could pop in an AMD 5x86-133 and run it with 4x40Mhz. That config was pretty popular for those who couldn't afford Pentiums (such as me). I ran some unknown PCI VGA and above Mach64 VT card with a 40Mhz PCI clock back then and never had any issues. Although such a comparison wouldn't make much practical sense.Your test clearly shows that all these cards roughly perform equally when being paired with a typical high performance CPU of the same time period.
Well I wouldn't say that my ISA tseng is faster than VLB, because my VLB cards all perform better as well. But the card is indeed getting pretty close to VLB cards in some benchmarks, unlike other ISA cards. Since you asked for a brand, I looked it up ... unfortunately, I didn't find anything, but I made some photos here : imgur.com/a/e68s1WB. I'll probably dump its bios, because maybe it'll work on other cards and would help speeding them up ? As for the turbo jumper, I think it simply removes some waitstates somewhere that mkake it run just as fast as possible.
Thanks for the photos. Being a 16 bit card on a 486, there should be a pile of wait states (it's about 3 on a 286, 10 on a 33MHz 386, and so surely many more on a 100MHz 486). It's not clear to me, but I think those are minimum values that the bus inserts for each byte accessed, so presumably the jumper brings it down to those values from a higher value. I'm not sure why they'd have it higher in the first place, but maybe so they can use different speeds of VRAM in the same PCB. But these are probably processor bus wait states, whereas I think the 0 wait state jumper on the card is probably for ISA bus wait states. So I don't know how much of a performance gain that would give, but I could imagine it coming close to explaining the difference. On the other hand, you ought to be able to remove the jumper and confirm that there is a big difference, right? Of course, in my video I merely meant that you were getting a higher framerate with this ISA card in your machine than I was getting with my VLB cards in my machine, not that it was faster than your VLB cards. It's quite clear your machine is also faster than mine. I don't have a ROM burner, so unless there is a flash utility, I probably wouldn't be able to use the ROM dump, but perhaps others will find it useful. As you are a RU-vidr, I could possibly loan you my TsengLabs ET4000AX card if you wanted to make a direct comparison in the same machine. If you ever get the urge to try it, let me know.
Looking into it some more, I see that some cards allow 0, 1, 2, 3 or even 4 wait states to be selected for writes to VGA ram. It's not at all clear to me what difference that would make, but people report about a 14-50% improvement on some cards, in some benchmark.
Your TsengLabs card seems identical to this United Solutions board: www.ebay.de/itm/United-Solutions-Ultra-Image-VGA-Board-Tseng-ET4000AX-ET4000-1MB-ISA/143397811362?hash=item21632cf4a2:g:Ey8AAOSwlh1dhmUI However, the FCC ID resolves to TsengLabs rather than United Solutions for your board, so perhaps it is some kind of reference design.
@@PCRetroTech very interesting ! I appreciate the proposition of loaning your card, but I don't plan to run benchmarks now and I haven't made any video in a while. But I'll keep that in mind if I need to study this. Or perhaps we could do this the other way around, but not right now, I am busy.
