With this many builds, I'm surprised that you are considering the same OS combination on almost all of them, even though in some case you intend to use the same chipset and CPUs just a few MHz apart.

That's also relevant for VGA choice, as wanting every build to do DOS and every build to do Windows means you need an all-rounder for all of them. It would be a lot more interesting if you make some builds DOS-only (where VESA SVGA support and raw framebuffer performance are key) and others Windows only (where high-res image quality and acceleration features make all the difference).

If you want do everything with everything, you will have to use all-rounders like S3 and CL chipsets (or Trident if you want a slow all-rounder). However ATi and Matrox chips are consummate Windows cards yet average to pretty bad at DOS, whereas ARK rules the DOS roost but a 1000PV is a bit anaemic for Windows (as is anything with less than 2MB for Win3.x and 4MB for Win95).

Note that S3 Virge cards are notorious for bad analog output (whitish/blueish haze over the whole image). Really good quality cards are fine (ELSA, Miro or Diamond, for example), but most generic ones are awful. Compare it side-by-side with something known-good (that Matrox MIllennium if you can find it) to see how good/bad it is.

I am kind of aiming for all round-ish builds at this point, they may not be optimised for everything. Trying to keep them semi-authentic for half year releases high end builds 1994-1996. This is what you might call a "tasting flight" of builds I want to do to feel the character of various notches of the performance continuum. Where putting highest benchmark parts together is more like sip and spit wine tasting, and a flight is more, "see how it goes down". Hoping to spend some days using each build, feeling it. Going on dates and experiences with them before settling down might be a good analogy. Since however they are machines, I could possibly swap a part or two around at the end. This is a small year limited segment of the hoard, which I want to thin out, so in the end, they might be competing with "ultimate" builds, dial a speed builds and even DOSbox and emulation to see where the "feel" wins out. I could benchmark 8 boards by 15 adapters and have to go through 120 combinations, but that'll just give me numbers not feel, and that doesn't even include every CPU I have that the board could take, so not likely to do it that way.

About the Tridents though I don't think the TGUI9680 is a slow card for it's era, these aren't like the ISA TVGA cards. The 9440 isn't all that bad considering it's a 1994 design, I think it's better than the GD5430 on PCI for example and the 9680s seem to keep up with the GD5436. Video RAM for windows, I am not into pushing the resolution envelope for these. Back in the day I was trying to run max resolution for webbrowsing, and it being one or two steps above "mainstream" was miserable for everything else. The UI designs tended to target common denominator resolutions and particularly for mass market stuff, they looked awful stretched out with elements too tiny and other problems. You never quite get the font right for all things either, you're always having to dicker with everything. Plus, windows fights you, the larger desktop you have, the more quickly you smash the GDI.EXE stack, so yeah for 3x/95 I am fine with 1MB... though again "ultimate" builds which span years may have different requirements. I feel like 800x600 is gonna be the standard on these, high color looks good enough, and a lot of windows games are gonna mess it up by locking back to 256 color anyway. Also these cards all have sockets and have one or two sets of RAM that could be tried out as required to bump some to 2MB.

Also they don't have to have "Perfect" DOS compatibility, just do a reasonable job on stuff +/- a year or two, I can probably have absolutely everything covered somewhere. I could have the P200 set up with a 19" CRT though and then spend hours struggling through Tomb Raider on an early P1 laptop with a slow LCD just because I'm amazed it could do it.

I guess it's hard to get a firm grasp of the requirements here, maybe something like "better than average at the time" (Apart from the pointless to get ahead of the curve for windows resolution thing) that you as a snotty pre-teen would have actually admired your parent for buying you, and didn't think they had cheaped out somewhere really essential, or had been completely clueless about something. Where we'll take the "time" as being a short while after release of the CPU at that speed, when it was actually available at retail.

BitWrangler wrote on 2024-02-26, 17:01:

I am kind of aiming for all round-ish builds at this point, they may not be optimised for everything. Trying to keep them semi-authentic for half year releases high end builds 1994-1996. This is what you might call a "tasting flight" of builds I want to do to feel the character of various notches of the performance continuum. Where putting highest benchmark parts together is more like sip and spit wine tasting, and a flight is more, "see how it goes down". Hoping to spend some days using each build, feeling it. Going on dates and experiences with them before settling down might be a good analogy. Since however they are machines, I could possibly swap a part or two around at the end. This is a small year limited segment of the hoard, which I want to thin out, so in the end, they might be competing with "ultimate" builds, dial a speed builds and even DOSbox and emulation to see where the "feel" wins out. I could benchmark 8 boards by 15 adapters and have to go through 120 combinations, but that'll just give me numbers not feel, and that doesn't even include every CPU I have that the board could take, so not likely to do it that way.

Still looks like doing virtually the same build time and time again - but hey, your builds, your choices.

About the Tridents though I don't think the TGUI9680 is a slow card for it's era, these aren't like the ISA TVGA cards. The 9440 isn't all that bad considering it's a 1994 design, I think it's better than the GD5430 on PCI for example and the 9680s seem to keep up with the GD5436.

Can't find any benchmarks comparing TGUI9680 with CD5436 or other later CL designs, but compared to anything else...

https://thandor.net/benchmarks/vga/vga?a=69&c=0&o=0&s=Submit
https://thandor.net/benchmarks/vga/vga?a=70&c=0&o=0&s=Submit

It comes in closer to Mach32 than Mach64 and actually slower than the TGUI9440 in Doom, and of the S3 options only the ancient 805 is slower. Then again, the difference vs top performers is only about 10%.

Video RAM for windows, I am not into pushing the resolution envelope for these. Back in the day I was trying to run max resolution for webbrowsing, and it being one or two steps above "mainstream" was miserable for everything else. The UI designs tended to target common denominator resolutions and particularly for mass market stuff, they looked awful stretched out with elements too tiny and other problems. You never quite get the font right for all things either, you're always having to dicker with everything. Plus, windows fights you, the larger desktop you have, the more quickly you smash the GDI.EXE stack, so yeah for 3x/95 I am fine with 1MB... though again "ultimate" builds which span years may have different requirements. I feel like 800x600 is gonna be the standard on these, high color looks good enough, and a lot of windows games are gonna mess it up by locking back to 256 color anyway. Also these cards all have sockets and have one or two sets of RAM that could be tried out as required to bump some to 2MB.

In Windows 3.x, 16b colour is more than good enough, but as someone who had to suffer 1MB and a monitor that could only do 800x600 without going eye-melting, I can assure you that 1024x768 is very, very desirable. 2MB gives you that at 16b colour, at least if the RAMDAC supports it.

Also they don't have to have "Perfect" DOS compatibility, just do a reasonable job on stuff +/- a year or two, I can probably have absolutely everything covered somewhere. I could have the P200 set up with a 19" CRT though and then spend hours struggling through Tomb Raider on an early P1 laptop with a slow LCD just because I'm amazed it could do it.

I guess it's hard to get a firm grasp of the requirements here, maybe something like "better than average at the time" (Apart from the pointless to get ahead of the curve for windows resolution thing) that you as a snotty pre-teen would have actually admired your parent for buying you, and didn't think they had cheaped out somewhere really essential, or had been completely clueless about something. Where we'll take the "time" as being a short while after release of the CPU at that speed, when it was actually available at retail.

Thing is, it's not one-dimensional, "better than average at the time" still has multiple things it could be better at. Analog output, supported resolutions and modes (sadly not always correlated with analog output), Windows acceleration performance, DOS SVGA VESA BIOS support, DOS scrolling, DOS framebuffer performance, acceleration of specific applications and indeed support for unusual operating systems or hardware platfoms (if you have an Alpha, you probably want an ATi video chip...).

But given what you say about actually wanting all the builds to bascially do the same, it does come down to allrounders I'm afraid.

Yes they are pretty much the same, but wanted to jump around between performance levels to avoid the "boiling a frog" effect, of slowly notching up performance on one board, and compare them on similar stuff, hell I might clone one multi OS install for all of them, though I expect the bottom to go back further and the top to go forward further, but more than 80% overlap incrementally. Though also, you put a spread of systems together for approx half years centered on any date and they're more or less the same. Even "big jumps" aren't big when you're looking at last of PIII vs first of P4 etc. Middle to middle of completed range is a lot different ofc. .... and believe it or not, this is the reasonable, abbreviated version... could double the number of boards and fill in every speed grade, but I never liked 50 and 60 FSB P1 CPUs so this is about finding what I like so I can cut most of that noise out right away, apart from P90 as it was a s5 "founder" as it were. It makes no sense at all unless you've got a whole stack of hardware you're looking to sort one way or another for curation into an actual collection that you cherish vs a whole stack of random hardware.

But graphics...

I am plowing through all the benchmarks I can find, just for idea of where cards sit, like that Thandor link. I was a bit uncomfortable trusting his single TGUI9680-1 data point, particularly as the -1 seems to be the low power cost reduced iteration... however, on looking at the two 9680s sitting here, that's what I've got on hand. Although I might have had another one I was using for second monitor on more modern gear that might be put away somewhere else. I probably sorted them for fastest for that so maybe that is a better/faster one, maybe it was one of these, IDK it's all a mystery sometimes. However, TheSerpentRider tests these cards too, and oddly they seem to do a lot better in the lower res VESA modes than ModeX, 640x480 and not too bad at 800x600, but fall off badly again at 1024.. So apparently not great for ModeX... but decentish on 640x480 VBE 2.0 .... Ergo, they might actually be better for the faster systems as the slower systems will not be trying 640x480 software render in anything, so fastest modeX cards would be preferable, while fast CPUs might be fine pushing a bit uphill on modeX .... we're still talking way above "90s playable" framerates at 30+ fps anyway so all good. Unless anyone can think of any ModeX games that a 166-200Mhz CPU actually struggles with at modeX software rendered, dipping below 20fps .... that aren't 1998 games that belong on a V2 or TNT2 and only have token software mode. That was previously specified, not looking to go more than a year or two upward.

There seems to be only two chipset/architectures with "consistent" performance over a lot of game and simulated gaming benchmarks, the Matrox ones always being near the top, and the GD5430 always being way out of the pack at the bottom, like half speed... damn I knew it was bad, but wow. Weird thing is I think the GD5429 is faster on VLB. A lot of other chipsets are all over the place... back of pack to front of pack... even between one modeX thing and another, or between two VBE tests.

I know CPU scaling is mostly a big thing for 3D with how heavy drivers are, and whether a game engine is heavy or light on it, but also not getting any idea of possible scaling effects on 2D PCI... Got the one extreme of the lower end Thandor results and the other extreme of high speed socket A SerpentRider results. I guess it's not driver based so much as BIOS code here. On ISA we know about some instances where 286 choke on cards with 386 optimised code in BIOS, do we know anything about 486-Pentium code optimisations??? Though also have CPU scaling considerations for cards that will just about do 3D... like Virge (original?) is supposed to be actually helpful at 133 but at higher speeds isn't, if DX was 40% better do we figure it's at least minor improvement or a wash on a P166 or what, is it worth using Matrox 3D on a P200?

So at the moment, I'm sorting the GD5430 into the "only in absolute desperation and the GD5434 ISA isn't possible" pile and the 9680s into the "depends, not modeX friendly for low CPU" pile, and the ARK1000PV into the "very modeX friendly" pile...

Also Mach64 is a weird one, a lot of people say they are overrated, and they don't give stellar results, merely okay on some things, but on other stuff, they are quite high, next to the Millenniums sometimes.

Aha, finding some better stuff with google vs what I was digging up on internal search...
486 results...
Re: Fastest PCI graphics card in a 486
and
PCI Graphics Roundup - DOS
Illustrate scaling better than anything else I have found. Not sure I can infer much from the ARK2000 results for the 1000, it seems more tweaked for higher res.
Redemption for the GD5430... if I was gonna run a PCI DX2-66 maybe it has a use 🤣

Sounds interesting, tedious but interesting.

dionb wrote on 2024-02-26, 22:03:

Then again, the difference vs top performers is only about 10%.

This is why I don't really worry about a cards performance on pure dos machines, the difference is almost never noticeable in actual game play.
Instead I'll choose a card on other factors like bus type, memory, compatibility, API support.

but it's still fun to read about different cards, the pro's the con's and how they perform.

Random find, newsgroup post with collection of Doom2 results from various systems.. https://groups.google.com/g/comp.sys.ibm.pc.g … n/c/JoURgxWOyK0