d'artagnan
10-01-05, 12:00 PM
I was reading through the Driver Heaven review and performance comparison of ATi's Crossfire platform when I cam across something that puzzled me. Namely the benchmark for a single 7800GTX on the CS:S Video Stress Test. (see here (http://www.driverheaven.net/reviews/crossfireatireviewxxx/source.htm)).
Their test rigs used an X2 3800 and they conducted their test at 1600*1200 with 4x anti-aliasing and 8x anistropic filtering and scored 119 fps. This struck me as seriously wrong as when I got my 7800GTX it was paired with an Athlon 64 3000+ and I immediately fired up the CS:S VST and tried it out at 1600*1200 with no anti-aliasing & bilinear filtering versus 6x anti-aliasing & 16x anisotropic filtering. The difference was less than one frame per second and was about 125 fps.
When I upgraded from my A64 3000+ to my A64 X2 4400+ I performed the same quick test and again saw less than a frame per second difference. Consequently I've always played it at 1600*1200 with 6x AA and 16x AF and had just assumed that I was still CPU limited.
Clearly something was up as an A64 3000+ just shouldn't be able to beat an A64 X2 3800+ - but was it a problem with their test rig & configuration or was it a problem with something else? So I endeavoured to do some testing of my own...
For these tests, my machine is an X2 4400+ on a Gigabyte K8NF9 nForce 4 mobo with a Gigabyte 7800GTX and one gig of PC3200 DDR RAM in a 4x256MB arrangement. The hard drives are on SATA-0 and SATA-1 and are 7200 RPM Maxtor DiamondPlus drives with 16MB of cache one of which is 120GB and the other is 300GB. The cooling is the stock cooling for all parts and the PSU is rated at 480W.
When running the CS:S Video Stress Test, all effects are at their maximum settings. With regards to core usage, all user processes other than steam.exe and hl2.exe are on core 0 whilst those two processes are on core 1. Let us begin...
1024*768
No AA, bilinear filtering - 158.53 fps
4x AA, 8x AF - 161.96 fps
6x AA, 16x AF - 159.96 fps
So, at 1024*768 there's no difference worth a damn. But we all knew this anyway.
1280*1024
No AA, bilinear filtering - [b]155.85 fps
4x AA, 8x AF - 140.20 fps
6x AA, 16x AF - 154.55 fps
What the ****? We see no tangible difference between no aa & bilinear filtering versus 6xAA & 16x anisotropic filtering although the frame rates are ever so slightly down compared to 1024*768 but... what is with that 4xAA & 8x AF score??? 4xAA & 8xAF is a fraction over ten percent slower than 6x AA & 16x AF.
1600*1200
No AA, bilinear filtering - [b]146.45 fps
4x AA, 8x AF - 121.97 fps
6x AA, 16x AF - 147.7 fps
Again, the scores for not bothering with AA and AF versus those with them maxed out are so samey that it could have been affected by something as slight as the way you looked at your monitor. To all intents and purposes, they're the same. However, the score for 4xAA & 8xAF is 17.5% lower than that for 6x AA & 16x AF.
This is wierd beyond belief. So, is it a problem with the Source engine, is it a curious quirk of my rig or is it a fundamental problem with running a 7800 at 4x AA & 8xAF? More worryingly, if it is a problem with running at 4x AA and 8x AF then is it one which scales upwards as you increase the resolution? If so, then not only could benchmarks being run at 1600*1200 with 4x AA & 8x AF be sorely understating the performance capabilities of the 7800 but it's a problem which would apparently get worse as you test on ever higher resolutions!
Gentlemen, there's only one way to be sure. You have to go forth and conduct your own tests to see if you can replicate this phenomenon. Not only that, you need to try out other benchmarks to see if they demonstrate the same problem. Get to it - your graphics card needs you!!!
p.s. for point of comparison, I'll include the scores when conducting the test so that all user processes (and I have quite a few running) have an affinity for both cores - all tests are conducted at 1600*1200.
No AA, bilinear filtering - 116.49 fps
4x AA, 8x AF - 104.10 fps
6x AA, 16x AF - 115.77 fps
Their test rigs used an X2 3800 and they conducted their test at 1600*1200 with 4x anti-aliasing and 8x anistropic filtering and scored 119 fps. This struck me as seriously wrong as when I got my 7800GTX it was paired with an Athlon 64 3000+ and I immediately fired up the CS:S VST and tried it out at 1600*1200 with no anti-aliasing & bilinear filtering versus 6x anti-aliasing & 16x anisotropic filtering. The difference was less than one frame per second and was about 125 fps.
When I upgraded from my A64 3000+ to my A64 X2 4400+ I performed the same quick test and again saw less than a frame per second difference. Consequently I've always played it at 1600*1200 with 6x AA and 16x AF and had just assumed that I was still CPU limited.
Clearly something was up as an A64 3000+ just shouldn't be able to beat an A64 X2 3800+ - but was it a problem with their test rig & configuration or was it a problem with something else? So I endeavoured to do some testing of my own...
For these tests, my machine is an X2 4400+ on a Gigabyte K8NF9 nForce 4 mobo with a Gigabyte 7800GTX and one gig of PC3200 DDR RAM in a 4x256MB arrangement. The hard drives are on SATA-0 and SATA-1 and are 7200 RPM Maxtor DiamondPlus drives with 16MB of cache one of which is 120GB and the other is 300GB. The cooling is the stock cooling for all parts and the PSU is rated at 480W.
When running the CS:S Video Stress Test, all effects are at their maximum settings. With regards to core usage, all user processes other than steam.exe and hl2.exe are on core 0 whilst those two processes are on core 1. Let us begin...
1024*768
No AA, bilinear filtering - 158.53 fps
4x AA, 8x AF - 161.96 fps
6x AA, 16x AF - 159.96 fps
So, at 1024*768 there's no difference worth a damn. But we all knew this anyway.
1280*1024
No AA, bilinear filtering - [b]155.85 fps
4x AA, 8x AF - 140.20 fps
6x AA, 16x AF - 154.55 fps
What the ****? We see no tangible difference between no aa & bilinear filtering versus 6xAA & 16x anisotropic filtering although the frame rates are ever so slightly down compared to 1024*768 but... what is with that 4xAA & 8x AF score??? 4xAA & 8xAF is a fraction over ten percent slower than 6x AA & 16x AF.
1600*1200
No AA, bilinear filtering - [b]146.45 fps
4x AA, 8x AF - 121.97 fps
6x AA, 16x AF - 147.7 fps
Again, the scores for not bothering with AA and AF versus those with them maxed out are so samey that it could have been affected by something as slight as the way you looked at your monitor. To all intents and purposes, they're the same. However, the score for 4xAA & 8xAF is 17.5% lower than that for 6x AA & 16x AF.
This is wierd beyond belief. So, is it a problem with the Source engine, is it a curious quirk of my rig or is it a fundamental problem with running a 7800 at 4x AA & 8xAF? More worryingly, if it is a problem with running at 4x AA and 8x AF then is it one which scales upwards as you increase the resolution? If so, then not only could benchmarks being run at 1600*1200 with 4x AA & 8x AF be sorely understating the performance capabilities of the 7800 but it's a problem which would apparently get worse as you test on ever higher resolutions!
Gentlemen, there's only one way to be sure. You have to go forth and conduct your own tests to see if you can replicate this phenomenon. Not only that, you need to try out other benchmarks to see if they demonstrate the same problem. Get to it - your graphics card needs you!!!
p.s. for point of comparison, I'll include the scores when conducting the test so that all user processes (and I have quite a few running) have an affinity for both cores - all tests are conducted at 1600*1200.
No AA, bilinear filtering - 116.49 fps
4x AA, 8x AF - 104.10 fps
6x AA, 16x AF - 115.77 fps