NVIDIA 3-Way SLI Technology Review – 8800 Ultra x 3!

Test System: Digital Storm in the House

As you might expect with the platform requirements as high as they are, NVIDIA was much more willing to send out complete systems for review rather than just a few cards in hopes that we had the rest of the items just laying around.  We DID have all the other components required to build an NVIDIA-recommend system for 3-Way SLI but NVIDIA still had a system integrator toss a box our way just in case. 


Digital Storm was tapped to provide us with a 3-Way SLI testing system and the machine they sent over is an absolutely gorgeous piece of computing power. Including a water cooled and overclocked Intel QX9650 processor, 2GB of Corsair DDR2 memory, an EVGA 680i motherboard and three EVGA 8800 Ultra graphics cards this computer would make just about any gamer excited enough to wet their proverbial pantaloons. 


We’ll be doing a full review of this system in its entirety next week, you can be sure of that, but for our 3-Way SLI performance testing today we yanked those GPUs out and installed them in our own GPU test bed that had slightly different components.  Our NVIDIA GPU test bed includes the same EVGA 680i SLI motherboard with three physical PCIe x16 slots and 2GB of Corsair memory but uses a dual-core Intel QX6800 processor running at 2.93 GHz.  To see if that quad-core CPU makes a difference for 3-Way SLI as NVIDIA hopes it does, we’ll be comparing the entire Digital Storm system to our own testbed in the system review next week. 

Our testing suite for this review is exactly the same as in recent GPU reviews – we are using the same hardware, same Vista x64 OS and same gaming titles.  Some of them will scale well and some of them probably will not but we wanted to take a look at 3-Way SLI from the gamers’ perspective; and gamers play a large variety of games.  Some of the titles we played with matched up with NVIDIA’s recommended games but not all of them did so we’ll see how it turns out as we go through our results. 

The results will show three different configurations of our 3-Way SLI system: one with three 8800 Ultras, another with two 8800 Ultras in the standard dual-card SLI configuration and finally another with just a single Ultra card.  This will allow us to see how the SLI performance scales from one to two to three cards and three times as much money.  Also, in most cases you’ll see the new 8800 GTS 512MB card thrown in there for a fourth data point and to add real-world, real-money comparison into the mix.

You’ll see that we have three sets of graphs in this review: one tested at 1600×1200, one at either 1920×1200 or 2048×1536 (depending on previous GPU test results) and one at 2560×1600.  The 25×16 tests will only include the 8800 Ultra cards as we didn’t test other GPUs at this level as most wouldn’t live up to the stress.  The only exception will be in Crysis – getting that at 2560×1600 just isn’t happening quite yet.

Testing Methodology

Graphics card testing has become the most hotly debated issue in the hardware enthusiast community recently.  Because of that, testing graphics cards has become a much more complicated process than it once was.  Before you might have been able to rely on the output of a few synthetic, automatic benchmarks to make your video card purchase, that is just no longer the case.  Video cards now cost up to $500 and we want to make sure that we are giving the reader as much information as we can to aid you in your purchasing decision.  We know we can’t run every game or find every bug and error, but we try to do what we can to aid you, our reader, and the community as a whole.

With that in mind, all the benchmarks that you will see in this review are from games that we bought off the shelves just like you.  Of these games, there are two different styles of benchmarks that need to be described.

The first is the “timedemo-style” of benchmark.  Many of you may be familiar with this style from games like Quake III; a “demo” is recorded in the game and a set number of frames are saved in a file for playback.  When playing back the demo, the game engine then renders the frames as quickly as possible, which is why you will often see the “timedemo-style” of benchmarks playing back the game much more quickly than you would ever play the game.  In our benchmarks, the FarCry tests were done in this matter: we recorded four custom demos and then played them back on each card at each different resolution and quality setting.  Why does this matter?  Because in these tests where timedemos are used, the line graphs that show the frame rate at each second, each card may not end at the same time precisely because one card is able to play it back faster than the other — less time passes and thus the FRAPs application gets slightly fewer frame rates to plot.  However, the peaks and valleys and overall performance of each card is still maintained and we can make a judged comparison of the frame rates and performance.

The second type of benchmark you’ll see in this article are manual run throughs of a portion of a game.  This is where we sit at the game with a mouse in one hand, a keyboard under the other, and play the game to get a benchmark score.  This benchmark method makes the graphs and data easy to read, but adds another level of difficulty to the reviewer — making the manual run throughs repeatable and accurate.  I think we’ve accomplished this by choosing a section of each game that provides us with a clear cut path. We take three readings of each card and setting, average the scores, and present those to you.  While this means the benchmarks are not exact to the most minute detail, they are damn close and practicing with this method for many days has made it clear to me that while this method is time consuming, it is definitely a viable option for games without timedemo support.

The second graph is a bar graph that tells you the average framerate, the maximum framerate, and the minimum framerate.  The minimum and average are important numbers here as we want the minimum to be high enough to not affect our gaming experience.  While it will be the decision of each individual gamer what is the lowest they will allow, comparing the Min FPS to the line graph and seeing how often this minimum occurs, should give you a good idea of what your gaming experience will be like with this game, and that video card on that resolution.

Our tests are completely based around the second type of benchmark method mentioned above — the manual run through.

System Setup and Comparisons  

Test System Setup
CPU	Intel Core 2 Extreme X6800 – Review
Motherboards	EVGA nForce 680i Motherboard – Review
Memory	Corsair TWIN2X2048-8500C4
Hard Drive	Western Digital Raptor 150 GB – Review
Sound Card	Sound Blaster Audigy 2 Value
Video Card	EVGA GeForce 8800 Ultra x 3 BFG GeForce 8800 GTS 512MB
Video Drivers	NVIDIA Forceware 169.18 Beta
Power Supply	PC Power and Cooling 1000 watt
DirectX Version	DX10 / DX9c
Operating System	Windows Vista Ultimate 64-bit

• Bioshock
• Company of Heroes
• Call of Duty 4
  
• Call of Juarez
• Lost Planet
• World in Conflict
• Unreal Tournament 3
• Crysis
• 3DMark06