While past reviews have shown that big multi-GPU performance gains are usually limited to more mainstream cards, today’s CrossFire results make a far better-than-expected impression.

AMD’s Radeon HD 6950 is anything but low-end, and yet its second card provides a huge 92% lead over a single board. Almost as impressively, a third card delivers an impressive 27% lead over two cards at “only” 2560x1600. We could expect even better results at higher resolutions, if only this lab had the monitors for that test. That’s why I just expensed two more 2560x1600 displays to support additional testing (just kidding, boss)!

Fears of three-way CrossFireX not being supported by a variety of games were almost put to rest when only a single game, Aliens vs. Predator, didn't realize any scaling improvements when we added a third card. While we still have minor concerns about what other games might suffer similarly (after all, AMD couldn't get F1 2010 working with anti-aliasing until zero-hour for our Radeon HD 6990 review), no performance gain in only one of five tested titles gives us a great deal of confidence in the company's compatibility efforts moving forward. Aliens vs. Predator is the one game where SLI showed its best scaling, a fact that we find odd given that this is not a “TWIMTBP” title.

CrossFire came out with a huge overall scaling lead over SLI, and removing the one title that didn’t reflect that average would have made the lead even bigger. Superior scaling allowed two mid-priced Radeon HD 6950s to approximate the performance of two higher-cost GeForce GTX 570s, while three HD 6950s took the performance win over three GTX 570s.

If CrossFire’s value-oriented victory in two-way and performance win in three-way configurations weren’t enough to convince us, the Radeon HD 6950's lower power certainly did. AMD started out with the highest efficiency at one card, and superb multi-GPU scaling allowed its efficiency to increase with every card we added. When’s the last time adding parts made your machine more efficient?

So, are two cards twice as powerful as one? Are three cards 50% better than two? We compared each two-way and three-way configuration to its own single-card scores to find out, and split the charts to remind readers that the GeForce GTX 570 SLI scaling analysis uses a GeForce GTX 570 baseline, while the Radeon HD 6950 CrossFire analysis uses a Radeon HD 6950 baseline.

Any Radeon HD 6950 CrossFire solution beats any GeForce GTX 570 SLI configuration at 2560x1600, regardless of whether two or three cards are used. This is probably the only resolution in today’s test that really matters, since a single card will pass most tests at anything less. Additionally, 2560x1600 has a similar pixel count as three 1280x1024 displays, so this is “just the beginning” for anyone planning to run in Eyefinity or Surround mode.

Things get ugly for Nvidia at 2560x1600. While a 67% performance gain for two-way and 112% gain for three-way SLI certainly aren't shabby, AMD’s 92% two-way and 245% 3-way gains put it in the winner’s circle.

This, in spite of the fact that one of our games, Aliens vs. Predator, didn’t even benefit from the third AMD card.

It takes three GeForce GTX 570s to match the scaling of two HD 6950s at 1920x1080. On the other hand, our previous page showed that Nvidia’s higher baseline made two-way configurations the only “proper” performance match.

Two HD 6950s are up to 78% better than one at 1680x1050, beating Nvidia’s 58% gain by a wide margin. A third GPU is nearly wasted by either technology at this setting, as we'd expect.

With Just Cause 2 (at our maximum test settings) making up the most-important portion of our power test, we used the corresponding benchmark values to calculate full-load efficiency. Average performance and power set the full-load baseline, while average power alone (for all six configurations) set the active idle baseline.

Oops! It might be acceptable for Nvidia’s better-performing card to consume more power in a single-GPU configuration, but CrossFire’s superior performance scaling makes the GeForce GTX 570 a poor efficiency choice in SLI. Note that the power was measured using a game as well as FurMark. This was essential, since the middle Radeon HD 6950 was completely idle during the FurMark test.

As I read reviews of the new 6990 I was amazed at the efficiency of the crazy 6990+6970 setup. Looks like AMD wants to be the multi-GPU champ, and hopefully Nvidia will prove ready to battle it out... I still prefer Nvidia hardware and drivers, but AMD's choice of putting large quantities of fast memory on their cards is clearly a paying one.

I still prefer Nvidia hardware and drivers, but AMD's choice of putting large quantities of fast memory on their cards is clearly a paying one.

My pair of GTX295 are really starting to struggle now with the new games, the fact NVidia had more efficiency multi-GPU drivers was the main reason I haven't considered going eyefinity yet... and to find out if anyone has ever had a issue turning a 6950 into a 6970.

This is what's great about competition: each side keeps trying to one-up the other and the consumer profits because the technology is constantly pushed forward. It's too bad this is lacking on the CPU side of things now. That and the fact that CPU engineers follow Moore's law as if it's scripture.