ATI Crossfire Multi GPU for the Radeons

How Crossfire Works

Supertiling

Supertiling is exclusive to ATI, Nvidia doesn't have a solution similar to this in their SLI technology. Supertiling shares the workload of a frame between the two graphics cards by splitting it up in the form of a chess board. Tile one is sent to the first graphics card for processing, tile two is sent to the second graphics card, then tile 3 is sent to the first card again and so on until the frame is fully rendered.

Supertiling sounds like a very complex method of splitting up a screen but take this as an example, the traditional method is to split the screen in half horizontally. The top half of the screen is sent to the first graphics card and the bottom half is sent to the second graphics card. Doesn't sound like a bad idea but what if the frame looked like the example below.

If the screen to the left was split horizontally in half and rendered by two separate graphics cards, then a small problem would have to be faced. The top half of the screen is mainly sky and doesn't have any complex textures to render., The bottom half of the screen has a floor with a complex texture and no single colour surfaces. The point here is that the first graphics card will complete its half of the frame well before the second. Then the first card will sit idle while the second card finishes the job, then the frame can be placed together and displayed on screen. Situations like this neglect the need for having 2 graphics cards as the performance increase is minimal.

Now lets take a look at the Supertiling example below. The frame is split into squares of 32 x 32 pixels and distributed from left to right, top to bottom in single tiles. As you can see, this way each of the graphics cards get a near equal amount of light and heavy work to do. This keeps the idle time down to a minimum and increases the efficiency and performance of the two cards working together.

As always though Supertiling isn't always the best way to go, this is why the crossfire system comes with 3 ways of splitting the workload. Supertiling does not support OpenGL and it doesn't give optimal performance all the time, only when the circumstances are right.

Scissor Frame Rendering (SFR)

Scissor frame rendering or SFR is very similar to the original method of SLI and Nvidia's Split Frame Rendering. SFR is the basic cut the screen in half and send half to the one graphics card and half to the other. There are some differences however between the original voodoo 2 SLI, Nvidia's Split Frame Rendering and ATI's Scissor Frame Rendering.

The original SLI simple cut the frame in half, no mathematics was required no calculations just cut in half and sent to two graphics cards. Nvidia's solution is the same but has a dynamic load calculation built in. It does not necessarily split the screen 50/50. It will calculate the load at the top of the frame and the bottom and split the frame accordingly.

ATI's solution is a little bit of both of the above. The screen is not always split 50/50 but its not calculated on the fly. This can save some calculations free up more clock cycles but can cause a slight drop from optimal efficiency. The load is set by the application / game and that value is fixed throughout the session. So if the screen is split 60 / 40 then that is how it will stay. The main advantage of the ATI solution here is that it can split the screen horizontally like the others, but also vertically. If splitting the screen vertically would give better results, then ATI's crossfire will win hands down. Nvidia doesn't currently have technology to work out the load split vertically down the frame.

>> Page 3 - Is it worth the expense >>

Back to Contents