NVIDIA 8800GTS 640 SLI vs. ATI HD 2900XT CrossFire - Testing explained
Author: Luka Rakamaric
Date: 26 Jul 2007

We won?t be going into details about R600 and G80 architectures, because those were covered in our previous articles. It will be enough to compare the specifications of the cards to understand the main differences. The 8800 has 96 shaders, while ATI declares it?s R600 to have 320 of them. Unfortunately for ATI, these two numbers are not directly comparable, but when you take a look at what each of the processors is capable of, and a little math, ATI?s number comes down to 64 processors (a number that is comparable to NVIDIA?s). Since the R600 has a little less than 250 MHz core clock over the G80 in 8800GTS, we can say that these are almost identical in terms of processing power. ATI on the other hand has a 512 bit memory bus, compared to the 320 bit of the G80, but it has ?only? 512 MB of memory, while the 8800GTS has 640 MB.

While testing CF and SLI setups, the problem that there?s no platform that supports both. We assembled two ?identical? test systems, with the exception of the motherboard. The SLI platform used our standard EVGA 680i SLI motherboard, and the CF rig had a DFI ICFX3200 T2R/G using ATI CrossFire Xpress 3200 chipset. The rest of the components were the same:
- OCZ Platinum XTC PC2-8000, 2x1024 MB, 4-4-4-12 @800MHz
- Intel Core 2 Duo E6700
- OCZ GameXstream 1010W
- Western Digital RaptorX 150 GB

One of the new things the latest generation of graphics cards brought to the market, aside from DX10 support, are the new AntiAliasing modes. NVIDIA introduced CSAA (coverage sample AA), and ATI indtroduced CFAA (custom filter AA). When combined with CF or SLI, some additional options pop out, such as 32XAA support in NVIDIA?s camp and 16xMSAA support for ATI. NVIDIA?s 8 series introduced a new AA method called CSAA, which is supposed to provide much improvement over the older MSAA method. 16xCSAA doesn?t provide the same quality as 16xMSAA, but it is quite close, and definitely above 8xMSAA. On the other hand, the performance penalty is somewhere between 4xMSAA and 8xMSAA.

ATI introduced CFAA, or custom filter AA, in which you are able to select what kind of sampling you would like to use. There are three settings that you can select. The first one is Box, which is the same as MSAA. The second is Narrow-Tent, which provides an ellipse that extends outside of the pixel boundaries from which the samples are taken. There is also a negative side to this. Because the samples are not within the pixel, there is a lot of blurring, especially with the next option, Wide-tent. Wide-tend extends the area the samples are taken from even further, so the blurring is quite visible.

All in all, we decided to test 4 AA methods for each setup.
For ATI:
- 16xMSAA
- 16xCFAA
- 8xCFAA
- 4xMSAA

For NVIDIA:
- 16xMSAA
- 16xCSAA
- 8xCSAA
- 4xMSAA

Unlike MSAA, CSAA and CFAA are not directly comparable, and in general CFAA produces better AA results but at the cost of quite noticeable blurring. This noticeable part may vary from user to user, but personally, I rate same level CSAA above the CFAA in terms of overall picture quality. This time around we did just DX9 benchmarks under WinXP, but we will update four DX10 benchmarks done under Vista as soon as possible.

Synthetic benchmarks used:
- Futuremark 3DMark06
- Futuremark 3DMark05

Games used:
- Company of Heroes 1.70 - built in benchmark
- F.E.A.R. 1.08 - built in benchmark
- Far Cry 1.4 - Regulator demo
- Oblivion 1.2.0416 - custom walkthrough with FRAPS
- Prey 1.3 ? custom timedemo
- Supreme Commander 1.0.3223 ? custom map
- X3 Demo ? run as benchmark option