a whole arsenal of graphic cards (nvidia, ati, kyro) coming in various flavors (gpu-model, agp/pci interface, gpu speed, memory amount, memory type, memory speed, memory bus width) is going to be sent through a benchsuite of 15 games and 2 synthetic bench programs in order to finally clear any doubt on what graphic card gives the best gaming performance on a k6/socket7 system. each card is run against each benchmark 6 times: the screen resolutions chosen are 1024x768, 1280x1024 (or 960) and 1600x1200, each resolution is run in 16bit and 32bit color mode. a complete run through the benchsuite produces 229 benchresults. we will see where each card shines most, at what color depth at what screen resolution in what game and where the card’s performance drops to an unplayable level. because of the vast amount of results we will also get an idea of what counts in a specific game situation and what not: gpu-speed, memory amount, memory speed etc. the test platform consists of an asus p5a, 1.03 motherboard, the cpu is a k6-3+-600. the system is equipped with 384 mb ram, the sound comes from a soundblaster 128pci. the operating system is win98se.
what this investigation is not about:
there is a lot of evidence (no, proof actually) that specific driver versions suit specific cards better than they do to others. and also it is just clever to find and use the best driver available. unfortunately due to time limitations it is impossible to test each card against every driver version available to find the driver that works best for it. furthermore driver A may give best results when the game is run at a certain resolution, but driver B is better in another resolution. the drivers I use for benching are not chosen deliberately but were taken after extensive research on the net and personal testing. there certainly may be better drivers for one or the other card. in case you know of and can provide actual data (except bench data resulting from testing with 3dmark or any other synthetic benchmark that is) then just let me know. so in short: this graphic card comparison is not primarily about finding the respective super-drivers but if we are lucky it may lead to this desirable result as a byproduct. also this thread is not about how to setup a game to produce highest frames per second. this would inevitably lead to lower visual settings (in-game settings, driver settings) but the game settings (driver settings) I use here are designed to please the gamer’s eye even at the speedier settings (lower screen resolution, lower color depth, etc).
____________________________________________________
update notification section:
every time I add a card’s results to the results table (below in extra post) I will update this very first post RIGHT HERE to inform you on significant news:
currently in the queue:
geforce 3 ti200 175/200
geforce 3 ti200 200/200
geforce 3
geforce 4 mx460
geforce 4 ti-4200
geforce fx 5200 pci
geforce fx 5700
geforce fx 5900xt
geforce 6200 pci
geforce 6600gt
results so far available:
15. July 07: Gainward GeForce 2 Ti500 Golden Sample, A-GF2Ti500-270/4/2-D64/250/128, a supercharged Geforce-2 card. due to the fact that driver 82.16 does not support gf2-series cards I had to use 71.84 instead.
5. July 07: nvidia gf2 gts pro, code A-GF2 GTS Pro-200/4/2-D64/182/128
26. June 07: Hercules Prophet II MX, code A-GF2MX-175/2/2-S32/183/128
26. June 07: nVidia GF2 MX400, code A-GF2MX400-200/2/2-S64/143/128
26. June 07: again nVidia GF2 MX400, code A-GF2MX400-200/2/2-S64/143/128
due to new driver issues that popped up (driver 44.03) I ran this card again with driver version 82.16 where possible. more details about that in another post.
graphic card’s code description:
A or /P: AGP or PCI; -GF/R/K: Geforce, Radeon or Kyro; -xxxx: model; -nnn: chip clock; /n: pixel pipes; /n: texture memory units; -S or D: SDR or DDR ram; /nnn: ram amount; /nnn: memory clock; /nnn: memory bus width (bits); note: DDR memory clock is denoted here with ist real speed, not effective speed (doubled).
_____________________________________________________
explanation of the layout of this thread – whole first page:
my approach is to put and collect all the vital information on the very first page of this thread. this helps tremendously (I hope) should the thread spread over many pages. one does not have to read through the whole thread to learn of the important information that would otherwise be spread deliberately and hard to find. so on the first page of this thread you will find the following information:
- how far it went so far (graphic cards done)
- what it is about
- the bench results (2 tables: absolute fps reached, comparison table)
- information on how to setup the respective game/bench/driver for those who want to do it on their systems; I will add information on request, just ask!
- other important information that may emerge
- a final resume
in detail: what is this thread about?:
through all the years all the while new posts appeared here (and at other forums as well) that asked, what would be the best (= fastest) graphic card solution for gameplay on a K6-system. although these threads received always a lot of attention and answers, nowhere I found a sound and systematic approach that resulted in some useful final answer. many people reported on hearsay or repeated what they have read somewhere but only a few took the task to support their allegations with data. what comes to my mind here are the preconceptions many folks have (and me too!) about agp versus pci, 64 bit vs 128 bit memory bus width, their brand addiction or aversion (for whatever reason) and alike. especially the notion that there is no benefit of using anything better than a gf2-mx card because a socket7 system is not powerful enough to max out even the lowly gf2-mx (which in fact it is not in the position to do) will be verified – or not. if people actually did provide numbers they missed some crucial points like that you can’t re-extrapolate what is valid on some other more advanced platform (athlon, p3, p4 etc) and that you can’t compare numbers taken from different s7-platforms (via, ali) where the testers setup their system with different os and different os-settings, use different cpu types at different speeds, varying fsb-speeds (overclocked agp-bus and ram), different amount of ram, different graphics driver versions and different driver setup, different sound solutions and, what is almost most important, use unaligned games/benchsoftware setups (ingame visuals settings). the only result you get from such comparisons is that you can say setup A produces higher fps than setup B, but you would never be able to tell whether graphic card A is faster than B.
so I decided many years ago to do it once and for all. I started to collect graphic cards and games over the time on ebay and I was quite successful to get most of them at very low prices (and sometimes not, but when you are on the hunt for something special you eventually loose control
). in the meantime I tried to find out how to tweak the ali-chipset (the basis of my setup) to give the video cards the best possible environment. I also did quite a lot of pretesting with some of the cards with different drivers to find the „one“ driver that fits all. although this is not the perfect approach to find the best graphic’s solution it is unfortunately the only way to go because of time restrictions. just imagine how much time would be needed to find the proper driver for each graphic card that suits it best if you are going to test 15+ cards and you have to pair each of them with about xx driver versions and send each of those combinations through the whole benchsuite. this is just impossible! and today – after I have finished the first three complete benchruns I can tell that I failed in finding such an „average superdriver“. I had to realize that the driver I had chosen so far (82.16) to run the nvidia cards with has serious trouble in some aspect and I had to use driver version 44.03 for that part of the bench suite (set 2) where the 82.16 failed. so I started the second card’s benchruns with driver 44.03 but also realized (since the two cards are quite similiar) that performance dropped that much in many single setups that I ran the second card again with driver 82.16 where applicable (means, those benches where the 82.16 does not fail – set 1). and as it appears now I will also have to add some driver version 56.xx for the gf-5xxx series because they are not fully (in all their flavours) supported by 44.03. again, this applies only to that part of the bench-suite where the 82.16 is in trouble. anyway I will go into more detail on this driver issue at a later time again here in this thread and/or eventually may start a whole new separate thread on it.setup of the test-machine:
asus p5a, rev. 1.03
my first socket7 board and I am still using p5a boards (or other ali-based boards) in all of my machines (except my test-rigs with via-boards and sis-board). very fast, rocksolid, easy to setup, what can you ask for more? well, it is not an overclocker’s dream ...
power supply:
JNC model LC-A300ATX
+3.3v: 20A, +5v: 30A, +12v: 10A, -12v: 0.8A, -5v: 0.5A, +5VSB: 2A
bios: 1011.005
external cache: disabled
pci/vga palette snoop: disabled
all rom bios shadowing: disabled
sdram config: 7ns
enhanced page mode count 16T
internal page detection: enabled
sdram pipe function: enabled
sdram x111-2111 mode : enabled
i/o recovery time: 2 busclk
graphics aperture size: 128mb
frame buffer posted write: enabled
agp bus turbo mode: enabled
k6 write allocate: enabled
passive release: enabled
delayed transaction: enabled
memory hole at 15-16m: disabled
cpu temperature reported: 49c
mb temp reported: 28c
pm os installed: yes
slot x irq settings: all auto
pci latency time 32 pci clocks
usb function: disabled
I deactivated onboard-usb in the bios in order to get agp gat-mode 2 running without stability problems. I learned from the speed boost by setting gat-mode to 2 many years ago but I always had trouble to reach a stable system. I just recently found out that turning onboard usb-device off does the trick. going from gat mode 1 to 2 gives 5 to 15 % speed boost depending on graphic card, game, screen resolution and color depth. so I am now able to set up the system like it was kind of „forbidden“ for many years because of stability issues. in fact, there are none anymore!
384 mb ram, 1x256 + 1x124, @100mhz, memory timing 2-2-2-4-7
K6-3+@600, @2.1v, 5.5x hw-multiplier
soundcard: sb128 pci ct4700
lancard: surecom 320xr
OS: win98se
swapfile set to min=max size of 256mb
„conservativeswapfileusage“-tweak
filecache set to min=max size of 16384kb
agp: driver 1.66 (very old indeed BUT according to a reliable source less restrictive in terms of speed-features deactivated compared to later versions; nevertheless solid as a rock even with quite „modern“ (in socket7 terms) cards)
nvidia driver 82.16 (from a cd that came with the pny geforce 6200 pci)
nvidia driver 44.03
ati driver: not determined yet
kyro driver: not determined yet
after booting into windows everest 2.20 reports (cpu 5.5x100):
memory read: 361 mb/s
memory write: 173 mb/s
memory latency: 182ns
changing cpu speed to 6x100:
memory read: 360 mb/s
memory write: 127 mb/s
memory latency: 182.6 ns
yes, a dramatic decline in memory write speed indeed!
after running any bench (just one), everest reports
read: 360mb/s
write: 137mb/s
latency: 188ns
here it is again, the boost in memory write performance reported by everest ..... precaching!, anyone?
no, it is not better to run the system @550mhz. the games need foremost cpu-power. memory speed comes next and it is memory read speed that counts more than write performance.
tools used:
- driver cleaner pro 1.5 (in case I have to remove the driver; helps to get rid of any trace of the old driver before installing a new one)
- rivatuner 2.0 final release (great and ABSOLUTELY important tool to setup and control the graphic’s driver)
- ctu: to setup the cpu for 600mhz, setup for mtrrs (only necessary for 2D vga write combining; not important here, but I do it anyway)
- wpcredit 1.4: to apply the tweaks for the ali-5 chipset
- mxk6opt 1.0: setting up the mtrr registers before windows starts (via call from autoexec.bat), helps ctu to find proper mtrr values;
- acronis trueimage 7.0: just in case of a desaster I can quickly rebuild the c-partition from a previous backup
- k6speed 0.88: just in case I would need it to activate write allocation; the p5a bios does it anyway.
- ali agp control center 1.4: manipulation of various agp driver settings including setting the gat-mode
games / benchmarks used:
I will send the graphic cards through the following games (all of them are full retail versions with the exception of dronez which is only the demo version, but is said to be equivalent to the real game):
open-gl based games:
- unreal tournament,
- quake 3 arena,
- quake 2,
- heavy metall fakk 2,
- half life,
- mdk 2,
- star wars jedi knight 2 outcast,
- star trek voyager: elite force,
- dronez,
- serious sam: first encounter,
- alice
d3d based games:
- expendable,
- mercedes benz truck racing,
- aquanox,
- max payne
the reasons why I selected these games for my benchmark suite are that they once were famous and well received in the gamers world and that they were used for many years as a de-facto toolsuite to measure graphics performance. so you will find tons of benchmark numbers on these titles all over the net to compare with. the big advantage to me as the test-conductor is, that there is plenty of information on the net on how to setup the games and all the dos and don’ts and finally there are instructions available on how to bench your system with these games.
some games I previously planned to include were: need for speed 5 porsche, undying, tomb raider angel of darkness, giants citizen kabuto and deus ex. they did not made it finally into the benchsuite because they either took very long to bench with (undying), are boring to watch (undying, giants), are of poor gameplay quality (giants), are a pain to bench (unreliable results due to driver cheats along nvidia and radeon cards eg. tomb raider), frequent freezes/crashes (nfs porsche) and finally I was not able to get a copy of the us-version of deus-ex which is necessary because the bench-scripts I found don’t work with the german version.
in the last minute I added 2 synthetic tests. first I wanted to stay away from them because whats really the use of it? does it tell you whether you can play a game at a decent speed? the answer is a clear NO! their results may be indicative in some sense and correlate somehow with real world games‘ experience but that is it. the reason why I in the end included
- glexcess (opengl)
- 3dmark 2001 (direct3d)
in my benchsuite is that glexcess gives us a lot of technical insight. it is far less cpu-power dependent than any game and displays the cards‘ capabilities. it splits its overall score into 4 more results that measure cpu performance, vram-speed, fillrate and polygon count separately and gives min-avg-max fps-scores for 11 different scenes respectively. if there is any tool that can show what a card could if we were not bound by the processor then this is it! glexcess is an open-gl based bench and I thought if I am going to include it I should also add some synthetic bench which is d3d – based. and voila, 3dmark 2001 made it finally into the bench-suite.
bench setup:
this was the toughest and most time-consuming part so far. I had a clear mind from the beginning that I wanted to give the lower-end cards a fair chance to show that they are able to run games decently. on the other hand I wanted to give the „high-end“ cards an environment where they can show their muscles. so I knew that I would run 2 sets, the first with 16 bit colors, the second with 32 bits colors. also I would use 3 screen resolutions which amounts to 6 bench-runs for each game (benchprogram):
screen res for 16 bit and 32 bit respectively:
1024x768
1280x1024 (or 960)
1600x1200
this part was quickly decided upon. the real work started with how to setup the game itself. in-game visual settings allow for a broad range of manipulation and can bring down a game to slide-show-speed. and then there are the driver’s settings, among them anti-aliasing AA and anistropic filtering AF which can be applied to each single frame.
if I now would setup each game‘s visual settings one for speed and one for eyecandy and combine it with the 6 settings above I would then have 12 setups to run through. if I then would pair each of them with AA-AF on / off I would end with 24 setups per game at minimum. I am mad enough to make a whole lot of crazy things (like doing this bench-thing on a k6-system in the year 2007!), but I am not a complete fool.
so I decided to make the 16bit runs the „speed“ runs and the 32bit runs the „quality“ runs.
speed runs are without AA and AF. the in-game settings are setup that way, that a geforce-2-mx gives at least 20 fps at 1024x768 (if possible). of course this varies from game to game and I had to find out how far I was allowed to crank up the visual settings (and sound settings too) or even lower them to make the gameplay smooth but also visually attracting.
quality runs are with AA 2x2 and AF 2x. the in-game settings are setup that way that a geforce-3-ti200 should give at least 24fps at 1280x1024 (or 960) @32 bit colors.
this way I stayed with 6 runs for each game/benchmark.
general driver setup rule (for nvidia cards; ati and kyro not played with yet):
before starting a game’s bench I setup the card’s driver via the context menue of the nvidia symbol in the tray and choose high performance / aa none / af none for the 16 bit speed setting and quality / aa 2x / af 2x for the 32 bit quality setting. the next step is to control and fine tune and sometimes even correct the driver’s setup via the tool rivatuner. here I make sure that the level of detail lod-setting is at 0 (d3d and opengl respectively, nvidia driver set up sometimes other values), that the aa/af is set accordingly (depending on what driver version you use, aa 2x may be set to 1,5x1,5 or something similar. correct this to 2x2 or 2x), that vsync is set to be always off, that dxt3 compression is used instead of dxt1. this has to be done only once as long as one does not open the nvidia driver panel again or restarts the computer. in other words: if you feel need to go into nvidia drivers settings, always check with rivatuner afterwards if everything is as wanted.
driver versions used:
as mentioned above I had to use 2 driver versions to run the benchsuite completely.
the following games/benches are run with driver 82.16 (set 1):
- unreal tournament,
- quake 3 arena,
- quake 2,
- heavy metall fakk 2,
- half life,
- mdk 2,
- star wars jedi knight 2 outcast,
- star trek voyager: elite force,
- dronez,
- serious sam: first encounter,
- alice
- glexcess
- expendable
for the follwing games/benches I use driver 44.03 (set 2):
- mercedes benz truck racing,
- aquanox,
- max payne
- 3dmark01 SE
in a future far far away:
once I am through with my collection of video cards I plan to pick the „best“ solution(s) and replace the K6-3+ by a K6-2+ to see how much impact the size of cpu on-die cache has. I also plan to run the benchsuite with the k6-3+ and k6-2+ respectively at 550mhz. maybe – if I am then still living – I will also pair the best card with a plain k6-2-550. but that is pure speculation.
also I plan to find the best sound solution. I again will take the best video solution(s) and let each of the different soundcards I collected over the time pass the whole benchsuite. this will comprise
- sblive value, sb audigy, sb audigy 2, hercules fortissimo II
I own an evga gf2-mx400 pci-based card. unfortunately this card does not run in my p5a. since the interface-connector shows only one notch I guess it is because the card is 3v only (pci 2.2) it works great on my asus p5s-vm (sis530) but also to my surprise with my shuttle hot 597 (via) board. if anyone has such a pci-card (gf2-mx) with two notches and would like to trade it for e.g.. a gf4-mx460 (agp) or a gf3-ti200 (agp) please tell me (pm). I really would like to include such a card here since it would make sense to compare it with my other gf2-mx cards (both agp). of course it is not useful to run this card on the sis or via board, since the results can not properly be compared with the p5a-results.
and yes, if it were not for the public I would not do this. I think a lot of people will follow this, the question this thread tries to answer is asked until today all over the forums (pc-related) I know of. it does not attract masses for sure but there is still enough interest in this field. for sure I will be a reviewer for the next months, not a gamer! 
agp speed is irrelevant when a pci video card is used 
