16-bit ISA VGA video card, circa 1988 386 16MHz
16-bit ISA VGA, circa 1991 486DX2-50
NVidia Riva 128 Pentium II 400MHz
Voodoo1 Pentium II 266MHz
nVidia TNT 2 Pentium 3 733MHz
ATI 7500 Pentium 4 1.4GHz, Athlon Thunderbird 1.2GHz
GF5 (5200 as well as faster models) Pentium II 233MHz
ATI Radeon 9800 Athlon XP 2300

leileilol, I wonder if I have worded my o.p. in an ambiguous way, and consequently it has been misinterpreted. Because looking at one of your "matches", the last one involving the ATI Radeon 9800 -> Athlon XP 2300, I think that if I were to replace the XP 2300 with an XP 2400, then the ATI Radeon 9800 would perform faster. Do you see what I mean?

rg100: I think there is a misunderstanding. If you have a certain graphics card the performance increases from a cpu limitation of the performance to a graphics card limitation of performance. Nethertheless if you increase CPU power even more the time where the cpu does calculations between frames gets still lower.
So as long as the card can receive already new data you will see still a performance increase.

Interesting is usually where the card reaches graphics card limitation thus the graph of performance increase shows a lower increase compared to the lower cpu power range.
I think leileilol wanted to point out at which cpu power this point is and I do agree with the estimation.
So fast 16 Bit ISA is about 486DX-40
GF2 GTS at a 1 GHz CPU f.e. Slot Athlon (GF4 MX440 is different, due to LMA2)
For Voodoo2 SLI I would estimate a 800 MHz Athlon, though it is more difficult because due to missing TnL those scale quite well with CPU power.
GF4Ti4600 maybe Athlon XP 3200+

I'm sorry, but I don't understand the wording used in your post. However, I do get the impression that I have made a mistake in my reply to leileilol. Sorry about that! In order to try to fully understand what is being said here, can I look at this example: GF5 (5200 as well as faster models) Pentium II 233MHz

So, if I remove the P2 @ 233MHz CPU, and replace it with a P2 @ 300MHz, I will not see any performance increase, if I run a 3D benchmarking program using a GF5 card? I'm afraid that's the only way that I am interpreting this thread, so far.

retro games 100 wrote:

I'm sorry, but I don't understand the wording used in your post. However, I do get the impression that I have made a mistake in my reply to leileilol. Sorry about that! In order to try to fully understand what is being said here, can I look at this example: GF5 (5200 as well as faster models) Pentium II 233MHz

So, if I remove the P2 @ 233MHz CPU, and replace it with a P2 @ 300MHz, I will not see any performance increase, if I run a 3D benchmarking program using a GF5 card? I'm afraid that's the only way that I am interpreting this thread, so far.

I think that remark from leileilol was a joke 😜

Anyway, interesting thread.

For me, I am more interested in the inverted case, whereby for a fixed CPU frequency, at what point does the next generation graphics card not show any performance increase?

If we were to fix the CPU speed to, say, a Pentium 233 MMX, and run thru the graphics benchmarks using these cards,

1) Matrox Millennium
2) Matrox Mystique
3) Matrox Millennium II
4) Matrox Millennium G200
5) Matrox Millennium G400
6) Matrox Millennium G450
7) Matrox Millennium G550
8) Matrox Millennium P650/P690
9) Matrox Millennium P750

at what point (at what card) would the benchmarks not increase in performance by more than 2%? Then iterate this process for various other CPUs like 133 MHz, 300 Mhz, 400 MHz, etc. This would let you know the "graphics card overkill point", whereas rg100 is wondering where the "CPU overkill point" is. More common terms might be "graphics limited" or "CPU limited", respectively.

If it helps, you can see things like this:

time taken to do something in a game (A) = time taken by processes not involving gpu (B) + time taken by processes involving gpu (C)

B will be largely dependent on cpu speed, so speeding up the cpu will reduce it, and thus reduce A.

C can also be affected by cpu speed, this is the part where the cpu can potentially bottleneck the gpu. Beyond a certain point, C is no longer affected by cpu speed, and it's gpu limited.

Both of these are subject to other factors [memory speed, hard disk access times, etc.] and also of course the way the game was written. Not sure if in old games there could even be an overlap between B and C tbh.

Hey RG100, I would simply go by slot type.

E.g. to test ISA cards I would take a P3 1.4 GHz or maybe an Ahtlon (some come with ISA slots apparently), for VLB I have no idea. I guess a fast 486?

PCI you can stick in any recent machine, as long as it runs the Operating system you want to use. For XP than can be a Phenom II or an i5/i7.

For AGP a Pentium 4 or Athlon 64 / FX or a S775 hybrid board (Asrock).

That way you wouldn't take any chances.

Actually, what I find the most interesting is the exact middle spot of where feipoa and RG100 are: At what point will adding a newer graphics card no longer help in improving framerates?
So to say, what's the most optimal graphics card to use with any particular CPU?

Hmm, I own a number of these and have tested them on a variety of platforms/ cpus. Roughly, I'd say:

Voodoo1 - P2 300mhz
Voodoo2 (SLI mode) - P3 600mhz
Voodoo3 up to 3500 - P3 600mhz
Matrox G400 - P3 600mhz
ATI 8500 - P4 2ghz
nvidia GF3 - P4 2ghz
GF4- P4 3ghz
GF5200 - P4 2ghz

sgt76 wrote:

Hmm, I own a number of these and have tested them on a variety of platforms/ cpus. Roughly, I'd say: […]
Show full quote

Hmm, I own a number of these and have tested them on a variety of platforms/ cpus. Roughly, I'd say:

Voodoo1 - P2 300mhz
Voodoo2 (SLI mode) - P3 600mhz
Voodoo3 up to 3500 - P3 600mhz
Matrox G400 - P3 600mhz
ATI 8500 - P4 2ghz
nvidia GF3 - P4 2ghz
GF4- P4 3ghz
GF5200 - P4 2ghz

Would this mean that the GF2 would be a good match for any P3-1Ghz?
I was thinking of matching these 2 together in the future

Thanks a lot. I thought about making a table with this data. How about this? I can edit it, when new data is suggested. When all of the rows in the table are complete, I can then resort it, based on the CPU speed. That way, you can get a general idea of how powerful these cards are, in relation to each other.

Edit 1: I have added a 3rd column, "System notes"

Just to clarify something here please. The idea is that I get to use all of my different mobos, from circa '88 to 2000+. Say I am working on a socket 7 mobo. What video card would be best suited for it? If a Voodoo3 can perform faster in a faster socket mobo (which it can, based on the data in the above table), then I wouldn't choose that video card for the socket 7 mobo. But say a Trio64V+ can't run any faster than a mobo using a 200MHz CPU, then I'd put a Trio64V+ in a socket 7 mobo, and use it with a Pentium 200 CPU.

I'm trying to think about what cards go in what boards. For a 16-bit ISA VGA card circa 1991, I now realise that I'm looking to put this card in a 486 board, and I need not bother looking for any CPU faster than about 40-50MHz. I can be satisfied that if I were to put in a faster CPU such as a DX2/66 CPU, this video card wouldn't give me faster frame rates, and so this CPU upgrade is not necessary.

Tetrium wrote:

Would this mean that the GF2 would be a good match for any P3-1Ghz?
I was thinking of matching these 2 together in the future

A GF2 would be a perfect match for a P3 1ghz.

Speaking of cpu bottlenecking, I've tested my GF3 and 4 cards on systems ranging from a P3 800mhz right up to a P4@ 3.6ghz. Just to give you a (very) rough idea of how much a P3 1ghz will hold back a GF3, with a P4 3ghz, NFS:MW is very playable at 800x600, low-medium details! Whereas a P3 1ghz with a 4600Ti plays GTA3 (much lower system requirements) or Far Cry (ditto) like arse.

For example, if I were to test a 3dfx Voodoo3 2000 card with various CPUs, at what point would it not matter if I used a faster CPU? This would enable me to understand the video card's speed limitation.

this would be highly dependent on what game/bench you are running, and what resolution/colordepth.
for me, i would always go for the fastest platform possible:
for all agp2x cards, use kt333+barton 3000+(166*13).
for all agp4x/8x cards, use pt880+pentium e6800(the boards can only run stable at around 1066fsb max, so higher clock ratio is important)

If you do a big testing / benchmarking marathon, I would definitely do charts, showing how the performance improves and then trails off. Would make for a very interesting study...

Having said that, it also depends heaps on the benchmark. One could show a graphics limitation (a faster CPU doesn't improve the score) like something with heaps of AA, another one a CPU limitation (score keeps going up with a faster CPU), like a 640 x 480 and very basic benchmark.

Guys, its not always CPU limitation only. Tbred and Bartons behave completely off the charts on different platforms - just compare VIA KT133A (great SDRAM chipset) vs NFORCE2 ultra. Doing 3dmark 2001 benchmark on Radeon 8500 on the same CPU (~2GHz) but different boards/memory is cca 8.7k marks vs 13k marks. Same improvement can be detected in Morrowind, I know because I tested this myself back in those days.

So I propose to include at least memory configuration too in the table.

Thanks a lot for the excellent suggestions. I have added a 3rd column to the table, called "System notes". If this can be improved, please let me know and I'll change it again. It's easy to do. Thanks.

I wouldn't think anything faster than a Virge would be anywhere close to fully utilized in a 486 PCI motherboard.

Partial utilisation is ok too. I'm at 1280x1024 on my 486 using a Matrox Millennium G200 PCI. If another graphics card increased Subspace performance by 10% or more, while not degrading office performance and clarity, I'd jump on it.

If I recall, my Virge DX didn't even come close to the clarity of my Matrox on this here 486. If I'm feeling really ambitious at some point, I'll run a few game frame rate tests at 1280x1024 with the two cards. Actually, I'm not sure a Virge DX w/4MB of RAM can handle that resolution, so nevermind.

If I had one of those SiS 486 motherboards, I'd like to see a showdown between the Matrox Millennium G200 and the Voodoo3 3000. While both cards are considered overkill for a 486, I'd like to see if one of them outperformed the other at 1280x1024, and if so, by what amount.