Quick fact: Shipped ATI chipsets in Q4/2005: Almost 7,000,000.

Benchmarks

Note that these results have been supplied by ATI.


In this picture we see a performance comparison between Dual x8 and Dual x16 PCI-Express lanes from both manufacturers' chipsets, ATI's new RD580 and NVIDIA's x16 SLI. Best performance with Dual CrossFire x16 is achieved in Serious Sam 2 where Dual CrossFire x16 is taking an advantage of more than 25% over Dual CrossFire x8. Performance gain is minimal on Splinter Cell: Chaos Theory, where Dual CrossFire x16 takes 3.67% gain over Dual CrossFire x8.

The best gains can be expected when there is a lot of PCI-E traffic going on. This is the case in bridge-less Crossfire setups like they are possible on the Radeon X1300 and Radeon X1600. During our own testing of the ASUS A8R32-MVP motherboard with two X1900 XTX cards in Crossfire there were no visible gains in 3DMark and FEAR. This is probably because all data the video card will ever need is uploaded to the on-board memory at level start. During actual gaming the video card will use the PCI-E bus scarcely so the additional bandwidth is not needed.
Performance gains will still be seen in applications which copy the renderdata to the system's memory and perform some calculations on it.
For example Lionhead has a Black & White 2 HDR + AA patch ready which enables High Dynamic Range Rendering for very nice visual effects but is still able to do Anti-Aliasing on these images. Every single image is post-processed by the CPU so massive PCI-Express bandwidth will be needed here.


Also the latencies are becoming important at this point. NVIDIA is using an additional chip on their x16 nForce chipset. This adds extra delays to the transfers because the second chip has to wait until the bus becomes available. While the data is copied, the GPU will sit idle and wait for the transfer to complete. This applies to any setup that runs in a multi-GPU environment. There is always some transfer of data to the card where the output device is connected to. It does not matter if the transfer is via bridge, display cable or PCI-E bus - the data has to be moved and it has to be synchronized. So what does it help if you have a world class video card that can render a lot of frames but every other frame it has to wait for the slow renderdata transfer to complete, effectively reducing performance.


Here we can see the data transfer penalty differences between ATI's and NVIDIA's chipsets. While the differences are rather small at current resolutions, testing at higher resolutions is a good indicator of what games to come will use on regular resolutions.


Performance comparison using ATI RD580 with single X1900 XTX and CrossFire. Best performance improvement is seen on Call of Duty 2 where Dual CrossFire x16 is almost twice as fast as the single card solution and the lowest improvement is on F.E.A.R, where the single card solution is slower about 25%.