View Full Version : AGP Aperture Size Test *VST*

Pages : [1] 2

12-30-04, 03:17 PM
I decided to do some testing with the different AGP Aperture sizes using Valve’s Video Stress Test (VST) for the benchmarking. The results from the test are to give a general ideal for the best AGP Aperture size. These tests were run twice, at my LCD’s native resolution 1600x1024, and a minimum IQ of 2xAA and 2xAF. The test were performed on both a BBA 9800 Pro & BFG 6800 Ultra OC with there latest drivers. Lastly my motherboard provides an AGP Aperture size for 512MB and 1GB, so it may not be present in most motherboard bios.

Results Key:
Red = First Run
Blue = Second Run

ATI 9800 256MB PRO:
1600x1024 2xAA 2xAF AGP-32MB 50.08fps/50.01fps
1600x1024 2xAA 2XAF AGP-64MB 48.85fps/48.85fps
1600x1024 2xAA 2xAF AGP-128MB 47.80fps/47.81fps
1600x1024 2xAA 2xAF AGP-256MB 35.70fps/35.68fps
1600x1024 2xAA 2xAF AGP-512MB 51.23fps/52.00fps
1600x1024 2xAA 2xAF AGP-1GB 51.02fps/51.08fps

BFG 6800 256MB Ultra OC:
1600x1024 2xAA 2xAF AGP-32MB 110.03fps/111.03fps
1600x1024 2xAA 2XAF AGP-64MB 89.25fps/87.08fps
1600x1024 2xAA 2xAF AGP-128MB 90.10fps/90.09fps
1600x1024 2xAA 2xAF AGP-256MB 70.30fps/71.06fps
1600x1024 2xAA 2xAF AGP-512MB 110.23fps/109.89fps
1600x1024 2xAA 2xAF AGP-1GB 109.26fps/110.02fps

It seems that the 256MB AGP Aperture size has the biggest drop-off in performance when running VST :retard:

I would like too see some of you guys/gals results :)

PS. The tests and both graphic cards where used in the FX rig (see sig).

This was also posted over at Rage3D (xmastong) (mikec)

12-30-04, 03:21 PM
Thanks! I can tell a lot of hard work went into this, and its greatly appreciated.

12-30-04, 03:23 PM
Thanks! I can tell a lot of hard work went into this, and its greatly appreciated.

Anytime (xmasmile)

12-30-04, 04:06 PM
wow... seems like the 32mb aperture size seems to be a nice one to go with :)

I'm using 512 personally @ the moment... will give it a decko and see what happens...


btw nice work :thumbsup:

12-30-04, 04:08 PM
Hm i dont get it how 32mb can be that good.

12-30-04, 04:22 PM
Bio - It'll be because the VST doesn't actually *need* any/much additional memory when working with a 256mb card. Therefore, when it's got a really small AGP aperture, it's forced to just fill up the on card memory rather than cleverly load balance for the chance of needing to load a lot into main V-Memory coming up. 'Least, that's what I'd guess.

12-30-04, 04:59 PM
That would seem like the plausible answer when using a smaller Aperture size. However, as you can see from my test, the jump from the 256MB Aperture size to 512MB Aperture size shows an increase (fps) rather than a decrease.

12-30-04, 05:03 PM
Strange... thanx for the benchies :)

12-30-04, 05:04 PM
OK...now I'm confused all to hell. So it's actually BETTER to have a lower AGP arpature in some cases? Wow...I'm gonna have to mess with this a bit and see what I should be running it at...I've been using 256megs because I didn't think it would make a difference...

Edit: oh my ****ING god...I just set it to 32megs, and I got a 400 point boost in 3dmark05 (almost 15% for me!). And most importantly, the parts of it that used to run EXTREMELY slow (like under 5 FPS) were suddenly running 2 to 3 times FASTER than they were with 256mb AGP Arpature. Unbelievable...I'm definatly going to be testing out a few settings, but basically I'm just getting back performance I should've had since the begining...I feel like I've been running in a marathon and just realised that I have weights on my ankles.

12-30-04, 05:22 PM
^^Same here bud^^

12-30-04, 05:23 PM
With modern cards setting the aperture size to 32Mb or less is like not using AGP at all (well, except for the faster data transfer). Looking at those results, I guess there's something weird going on with the drivers and/or Source engine. The framerate drop may be due to the card doing AGP-texturing instead of caching textures in the onboard RAM.

Which is something to be avoided completely. Since the early days of AGP, GPUs like the Intel740 or the S3 Savage demonstrated that unecessary use of the AGP bus leads to performance penalities.

12-30-04, 05:30 PM
I will test 3dmark05.

12-30-04, 05:35 PM
Any chance you could run the same tests (maybe fewer of them to make it easier) on different games? Maybe a Half-Life 2 demo (non VST) or maybe Doom 3 or something like that. These results are amazing to say the least. All I can say is, I sure hope im running at 256 and not 512. I want a speed boost :D

12-30-04, 05:39 PM

12-30-04, 05:42 PM
OK, I just did some original testing:

I set my AGP aperture to 8mb, and 3dmark didn't even work (just went to a black screen after loading, I restarted the computer after nothing displayed at all). However, I set it to 16megs, and I got even BETTER overall performance in Game Test 1 in 3dmark05 than I did with 32megs. BUT (and this is a big but) Game Test 2, 3, and parts of 1 ran EXTREMELY bad in some areas (1-2 FPS) where you could see a large part of the level, i.e. where the most textures were on the screen. So, here's my current theory: the less your AGP Arpature size, the more textures will be stored directly on the video card's RAM and the faster it will render scenes, but if it uses more texture data than what's availible on your videocard RAM or your AGP Arpature size, your performance get thrown out the window while it apperently loads the textures from the harddrive...or something? Anyway, it seems the best setting for most programs would probably be 32mb, however I think having a higher AGP Arpature would benefit games that have LOTS of texture data in a single scene (i.e. Doom 3 in Ultra-high quality mode). I'll have to test with Farcry to see if my theory holds any water.

Oh, and I have a 128meg 6800oc. If you have a 256meg card, your results may differ (hmm, wonder if 256meg cards could get away with running an 8mb or 16mb aperture...)

12-30-04, 05:58 PM
We must look into this. Looks very interesting.

12-30-04, 06:03 PM
Holy crap, I just tried playing Doom 3 with a 32meg aperture, and it's UNBELIEVABLY smoother. In both the last boss fight and the Guardian fight I'm easilly getting DOUBLE the fps that I used to, and this issue I was having (http://www.nvnews.net/vbulletin/showthread.php?t=42652) is COMPLETELY non-existant now. I KNEW my performance was far worse than it should've been, I just KNEW it. Thank you topic creator, you've saved my life (dozens of times in fact, now that my framerate is so much better I'll actually be able to aim at stuff :D ).

12-30-04, 06:06 PM
Well, I hate to say it but it did nothing for me :(

1600x1200 8xAF no AA (AA\AF set in game, control panel set to app selection), all in-game settings maxed, vsync off, high-quality filtering set in control panel.

81.29fps at 256Mb (had been running at this for a long time)
81.49fps at 512Mb
81.45fps at 32mb

My BIOS doesnt go lower than 32 or higher than 512. So thats that. For what its worth, im using 67.02 drivers. BTW, the first part of the test, when its looking at the various pixel shaded blocks (i think its the 3rd one specifically) it drops to 33fps exactly. Other parts of the test run at 180. Is this normal for this part of the test?

12-30-04, 07:17 PM
I've had the opposite and somewhat expected effect (sorry, no benches at present). When I set up this machine (1gm ram, FX5800) I hadn't yet configured AGP which was set at 64mb. 3DMark2003 and Doom3 ran okay, but somewhat slow, increasing to 128 then 256mb caused the framerate to increase by a significant amount perhaps 20-30%. I didn't try higher than that. What people said about 32mb effectively not using AGP is true, the actual allocatable memory used by AGP is something like (size - 16) / 2 (eg. adjusted for some fixed amount and double buffering.) It would be good to repeat the AGP test with a variety of games/benchmarks and video cards. I would personally not recommend lowering AGP aperture based on this one benchmark (although the results were most likely accurate for that test).

12-30-04, 07:38 PM
Using my 9800 Pro with the same settings listed above in my first post. Playing HL2 with the varying AGP Aperture sizes, it seems the most you’ll receive is about 3-4fps gain with the 128MB setting. However it does conclude the 256MB AGP Aperture isn’t the greatest to be using. Also I included some capture shot’s of a particular area to show some difference within the varying apertures frame rate. If you look at the 256MB AGP Aperture pic, you can see that the frame rate is lower, yet it has less going on within the shot.

ATI 9800 256MB PRO:
1600x1024 2xAA 2xAF 32MB 61.40fps/61.40fps
1600x1024 2xAA 2XAF 64MB 60.87fps/61.02fps
1600x1024 2xAA 2xAF 128MB 63.02fps/65.11fps
1600x1024 2xAA 2xAF 256MB 59.17fps/60.00fps
1600x1024 2xAA 2xAF 512MB 61.02fps/60.78fps






PS. Since the 9800 Pro was already in the computer (last to be used during my first test). The FPS would look even better at the varying AGP size on my 6800 card (xmasmile)

12-30-04, 09:09 PM
Ugh, this is getting annoying. At 16 meg aperture, Gametest 1 in 3dmark2005 runs beautifully and is almost always running at over 15FPS, except for two specific areas where it runs at less than 1 FPS, and half of Gametest 2 also runs at 1 FPS. With 32meg aperture, overall performance is much lower: areas that ran at over 20 FPS with 16 meg aperture ran at less than 10 FPS with a 32 meg aperture, but all of the Gametests didn't have any problem with dropping to 1 FPS like they did with a 16 meg aperture. And game performance is similarly screwy: in Doom 3 some areas (like the previously mentioned last boss fight and when close to a disintigrating body) ran MUCH better with a 32 megabyte aperture than they do with a larger aperture, but with a 32 meg aperture many of the bigger areas in Half-life 2 run at less than 5 FPS, while running it at a higher aperture completely eliminates this problem. Both games were totally unplayable at 16 megabyte aperture.

So basically: low aperture is best for games that use GPU-heavy effects but minimal textures, while a high aperture is basically REQUIRED for games that use a lot of texturing. Ugh, This is getting rediculous...I should NOT have to choose which games I want to run well at any specific time, nor should I have to make a tradeoff between getting much worse overall performance or areas where my framerate drops to unplayable levels. Jeez, there HAS to be a way to fix this and have better performance WITHOUT the huge framerate drops, I honestly don't think I can put up with changing my aperture every single time I want to play a different game. Oh, and here's my general performance tests I got from running 3dmark if anyone is interested. Just remember that in the 16meg tests, GT1 and GT2 got massive framerate drops in certain areas (more in GT2 than in GT1).

XP2000, KT266a mobo, 768mb PC2100 RAM, BFG 6800oc clocked at 400/820

--16meg aperture--
3dmarks: 2830
GT1: 14.7
GT2: 4.6
GT3: 21.3

--32meg aperture--
3dmarks: 3581
GT1: 12.1
GT2: 11.2
GT3: 21.5

--64meg aperture--
3dmarks: 3475
GT1: 10.8
GT2: 11.5
GT3: 21.6

--128meg aperture--
3dmarks: 3451
GT1: 10.6
GT2: 11.5
GT3: 21.5

12-30-04, 09:28 PM
Hmm....I wonder if theres a PCI-Express equivalent to this? I don't see anything like it in my bios.

12-30-04, 09:29 PM
I pretty much came to the conclusion that any game using vast amounts of high quality texture data (Far Cry, Doom 3, HL2, ect..) should use 128MB – 512MB AGP Aperture setting. The 128MB setting being the best in my opinion (xmasmile) .

While games like GTA:VC, Blood Rayne, and earlier Unreal based engines get a better performance with the lower AGP Aperture settings.

Nitz Walsh
12-30-04, 09:34 PM
I pretty much came to the conclusion that any game using vast amounts of high quality texture data (Far Cry, Doom 3, HL2, ect..) should use 128MB – 512MB AGP Aperture setting. The 128MB setting being the best in my opinion (xmasmile) .

While games like GTA:VC, Blood Rayne, and earlier Unreal based engines get a better performance with the lower AGP Aperture settings.
This is very interesting - perhaps one more reason to hold off on purchasing a $400 6800 card until its PCI-E native. :)

Thanks to the OP and Edge for doing additional testing, as if I did purchase a 256meg 6800, first thing I would have done is set the AGP aperture to 256 meg.

It would be interesting to see ATI/Nvidia's comments on this...

12-30-04, 09:39 PM
It would be interesting to see ATI/Nvidia's comments on this...

Their to busy plotting the future of PCI Express cards (mikec)