Testbed and Methods

We are going to investigate the performance of multi-GPU graphics systems in contemporary games using the following testbed:

• Intel Core i7-965 Extreme Edition processor (3.2GHz, 6.4 GT/s QPI);  
• Gigabyte GA-EX58-Extreme mainboard (Intel X58 Express chipset);
• Corsair XMS3-12800C9 (3x2GB, 1333MHz, 9-9-9-24, 2T);
• Maxtor MaXLine III 7B250S0 HDD (250GB, Serial ATA-150, 16MB buffer);
• Enermax Galaxy DXX EGX1000EWL 1000W power supply;
• Dell 3007WFP monitor (30", 2560x1600@60Hz max display resolution);
• Microsoft Windows Vista Ultimate SP1 64-bit;
• ATI Catalyst 9.5 for ATI Radeon HD;
• Nvidia GeForce 185.85 WHQL for Nvidia GeForce.

The graphics card drivers were configured in the following way:

Nvidia GeForce:

• Texture filtering – Quality: High quality
• Texture filtering – Trilinear optimization: Off
• Texture filtering – Anisotropic sample optimization: Off
• Threaded optimization: Auto
• Vertical sync: Force off
• Antialiasing - Gamma correction: On
• Multi-GPU performance mode: NVIDIA recommended
• Multi-display mixed-GPU acceleration: Multiple display performance mode
• Set PhysX GPU acceleration: Enabled
• Ambient Occlusion: Off
• Other settings: default

ATI Catalyst:

• Smoothvision HD: Anti-Aliasing: Use application settings/Box Filter
• Catalyst A.I.: Standard
• Mipmap Detail Level: High Quality
• Wait for vertical refresh: Always Off
• Enable Adaptive Anti-Aliasing: On/Quality
• Other settings: default

The list of benchmarks includes the following gaming titles and synthetic tests:

First-Person 3D Shooters

• Call of Duty: World at War
• Crysis Warhead
• Enemy Territory: Quake Wars
• Far Cry 2
• F.E.A.R 2: Project Origin
• Left 4 Dead
• S.T.A.L.K.E.R.: Clear Sky

Third-Person 3D Shooters

• Devil May Cry 4
• Prince of Persia

RPG

• Fallout 3
• Mass Effect

Simulators

• Race Driver: GRID
• Tom Clancy's H.A.W.X.

Strategies

• Red Alert 3: Uprising
• World in Conflict: Soviet Assault

Semi-synthetic Benchmarks

• Futuremark 3DMark06
• Futuremark 3DMark Vantage

We selected the highest possible level of detail in each game using standard tools provided by the game itself from the gaming menu. The games configuration files weren’t modified in any way, because the ordinary user doesn’t have to know how to do it. We made a few exceptions for selected games if that was necessary. We are going to specifically dwell on each exception like that later on in our article.

The following graphics accelerators and multi-GPU configurations participated in our today’s test session:

• ATI Radeon HD 4890 3-way CrossFireX (3xRV790, 850/850/3900MHz, 2400sp, 120tmu, 48rbe, 3x256-bit, 3x1024MB GDDR5)
• ATI Radeon HD 4890 CrossFireX (2xRV790, 850/850/3900MHz, 1600sp, 80tmu, 32rbe, 2x256-bit, 2x1024MB GDDR5)
• ATI Radeon HD 4890 (RV790, 850/850/3900MHz, 800sp, 40tmu, 16rbe, 256-bit, 1024MB GDDR5)
• ATI Radeon HD 4870 X2 4-way CrossFireX (4xRV770, 750/750/3600MHz, 3200sp, 160tmu, 64rbe, 4x256-bit, 4x1024MB GDDR5)
• ATI Radeon HD 4870 3-way CrossFireX (3xRV770, 750/750/3600MHz, 2400sp, 120tmu, 48rbe, 3x256-bit, 3x1024MB GDDR5)
• ATI Radeon HD 4870 X2 (2xRV770, 750/750/3600MHz, 1600sp, 80tmu, 32rbe, 2x256-bit, 2x1024MB GDDR5)
• Nvidia GeForce GTX 285 3-way SLI (3xG200b, 648/1476/2484MHz, 720sp, 240tmu, 96rbe, 3x512-bit, 3x1024MB GDDR3)
• Nvidia GeForce GTX 285 SLI (2xG200b, 648/1476/2484MHz, 480sp, 160tmu, 64rbe, 2x512-bit, 2x1024MB GDDR3)
• Nvidia GeForce GTX 295 (2xG200b, 576/1242/1998MHz, 480sp, 160tmu, 56rbe, 2x448-bit, 2x896 MB GDDR3)
• Nvidia GeForce GTX 285 (G200b, 648/1476/2484MHz, 240sp, 80tmu, 32rbe, 512-bit, 1024MB GDDR3)

We ran our tests in the following resolutions: 1280x1024, 1680x1050, 1920x1200 and 2560x1600. Everywhere, where it was possible we added MSAA 4x antialiasing to the standard anisotropic filtering 16x. We enabled antialiasing from the game’s menu. If this was not possible, we forced them using the appropriate driver settings of ATI Catalyst and Nvidia GeForce drivers.

Performance was measured with the games’ own tools and the original demos were recorded if possible. We measured not only the average speed, but also the minimum speed of the cards where possible. Otherwise, the performance was measured manually with Fraps utility version 2.9.8. In the latter case we ran the test three times and took the average of the three for the performance charts.