SLI vs. CrossFire, Part 1 – mid-range multi-GPU scaling & value
Test Configuration
Test Configuration – Hardware
- Intel Core i7 920-reference 2.66 GHz and overclocked to 3.8 GHz; 21x multiplier for 3.97 GHz, Turbo is on.
- Gigabyte EX58-UD3R (Intel X58 chipset, latest BIOS, PCIe 2.0 specification; CrossFire/SLI 16x+16x).
- 6 GB OCZ DDR3 PC1800 Kingston RAM (3×2 GB, tri-channel at PC1600 speeds; 2×2 GB supplied by Kingston)
- Galaxy GTX 560, 1 GB custom cooler and clocks, Galaxy Clocks 835/2004 MHz and stock-clocked to 822/2004 MHz, supplied by Galaxy under NDA.
- GeForce GTX 560, 1 GB reference design and clocks (822/2004 MHz) supplied by Nvidia under NDA.
- GeForce GTX 580, 1.5 GB reference design and clocks (772/2004 MHz), supplied by Nvidia
- GeForce GTX 480, 1.5 GB reference design and clocks (700/1401 MHz), supplied by Nvidia
- GeForce GTX 570, 1.2 GB reference design and clocks (732/1900 MHz), supplied by Nvidia and further overclocked to 815/2000 MHz.
- EVGA GTX 460 FTW; 1 GB, overclocked version (at reference clocks 675/1800 MHz) supplied by Nvidia/EVGA
- Galaxy GTX 460 SOC; 1GB overclocked version at reference clocks (675/1800 MHz), supplied by Galaxy
- Galaxy GTS 450 SOC; 1GB overclocked version at reference clocks (783/1804 MHz), supplied by Galaxy
- GeForce GTS 450, 1 GB reference design and clocks (783/1804 MHz), supplied by Nvidia
- ATI Radeon HD 6970 (2GB, reference clocks, 880/1370 MHz) supplied by AMD
- ATI Radeon HD 6950 (2GB, reference clocks, 800/1250 MHz; also flashed to stock HD 6970) supplied by AMD
- ATI Radeon HD 5870 (1GB 850/1200 MHz) by Diamond
- ATI Radeon HD 5870 (1GB 850/1200 MHz) supplied by AMD/Powercolor
- ATi Radeon HD 6870 (1GB, reference clocks, 900/1050 MHz) supplied by AMD
- ATi Radeon HD 6870 (1GB, reference clocks, 900/1050 MHz) supplied by AMD
- Onboard Realtek Audio
- Two identical 500 GB Seagate Barracuda 7200.12 hard drives configured and set up identically from drive image; one partition for Nvidia GeForce drivers and one for ATI Catalyst drivers
- Thermaltake ToughPower 775 W power supply unit supplied by Thermaltake
- Thermaltake Element G Case supplied by Thermaltake
- Noctua NH-U12P SE2 CPU cooler, supplied by Noctua
- Philips DVD SATA writer
- HP LP3065 2560×1600 thirty inch LCD
Test Configuration – Software
- ATi Catalyst 11.1a performance (beta) driver for all Radeons; highest quality mip-mapping set in the driver; surface performance optimizations are off; “use applications settings” are checked
- NVIDIA GeForce release candidate 266.66 for GTX 560 Ti; WHQL 266.58 used for GTX 580, GTX 570/GTX 480/GTX 560. High Quality
- Windows 7 64-bit; very latest updates
- DirectX July/November 2010
- All games are patched to their latest versions.
- vsync is forced off in the control panel.
- Varying AA enabled as noted in games and “forced” in Catalyst Control Center for UT3 ; all in-game settings are specified with 16xAF always applied; 16xAF forced in control panel for Crysis.
- All results show average, minimum and maximum frame rates except as noted.
- Highest quality sound (stereo) used in all games.
- Windows 7 64, all DX9 titles were run under DX9 render paths, DX10 titles were run under DX10 render paths and DX11 titles under DX11 render paths.
The Benchmarks
- Vantage
- 3DMark11
- F.E.A.R.
- X3:Terran Conflict
- Enemy Territory: Quake Wars
- Call of Duty 4
- Unreal Tournament 3
- Batman: Arkham Asylum
- Grand Theft Auto IV
- Serious Sam, Second Encounter HD (2010)
- Wolfenstein
- Left 4 Dead
- Grand Theft Auto IV
- Borderlands
- Mafia II
- Call of Juarez
- Crysis
- Warhead
- Lost Planet
- World in Conflict
- Far Cry 2
- Just Cause 2
- H.A.W.X.
- Resident Evil 5
- Alien vs. Predator
- Battleforge
- STALKER, Call of Pripyat
- Dirt 2
- F1 2010
- Metro 2033
- Lost Planet 2
- H.A.W.X. 2
- Heaven 2
We have got an interesting project going. Let’s check our results.
I’m not 100% certain, but to analyze microstuttering, place a check in the box next to “Frametimes” in Fraps. Then when you press the hotkey, it will create a log file with a timestamp when each single frame was outputted. Only a few seconds is enough to make the log file really, really long. Then take a portion out of the log file and make a chart out of it, that measures the time between each timestamp, to see if the frames are consistent with each other in similar intervals, or if every other frame is too close to the other one.
If a game runs at say, 45fps with your SLI or CF setup, but feels more like 23-30fps, then definitely analyze this with FRAPS.
Great review so far.
How do the numbers change, if at all, if Split Frame Rendering is used instead of Alternate Frame Rendering?
The last time I used SLI was with my Voodoo2 3000s. It was a gigantic waste of $200, in 1996 dollars.
If SFR eliminates micro-stutter without too much of a performance penalty I might have to try SLI again.
why don’t they add BF:BC2?
and also 6950 n 6970 crossfire?
Concerning the microstutter, frames time (using that fraps option) is supposed to fluctuate more erratically on crossfire/sli than what it would be on a single card. I think instead of testing a moving scene, it would make more sense to test it on a completely still scene for a few seconds and see how they compare in the excel output file. You don’t want a moving scene because then you won’t be able to differentiate between the erracticness you would get from a moving scene and the erraticness you would get from microstutter.
Another interest option would be to downclock a sli/crossfire setup to a point where it matches the average framerate of the single card. This way you could could see if the multi-gpu setup looks choppier than a single card despite having the same average frame rate.
Excellent work! At the end, simple recommendations would have been nice. =)
Please include Civilization 5 if possible the next time you benchmark.
It is an important game which will test the tesselation feature and its scaling ability in multi-gpu configurations.
Civilization 5 has been added to my benching suite along with DiRT 3 and Total War, Shogun 2.
You’ve done a great job of benchmarking gaming performance, but including charts with FPS vs $$, and $$ vs wattage would be much more useful.
The wattage (both idle and load) figures can be especially important, as some of these cards can easily draw more juice than all but the most powerful (and expensive) power supplies can provide — and that definitely factors into the cost analysis.