CrossFire Rendering Modes

ATI has three different types of rendering modes with CrossFire. The first is shared with NVIDIA's SLI implementation and that is Alternate Frame Rendering (AFR). AFR is quite simple, the cards render every other frame; the first card renders the even frames while the second card renders the odd. ATI claims that this mode offers the highest performance increase out of the three supported rendering modes as the other two cannot take full advantage of the vertex processing on both cards. AFR does not work in games where the successive frame is dependant on the current frame as frames on each GPU are generated independently. In this case, ATI falls back onto the next rendering mode, Supertiling.

Supertiling is exclusive to ATI's CrossFire and NVIDIA does not have an equivalent mode with SLI. With Supertiling the screen is divided up into 32x32 pixels squares like a chess board and the two cards render alternate tiles. ATI's explanation is that the work load is more efficiently divided up with the supertiling method as complex objects are more evenly split between the two cards. The example shown in ATI's documents is a pair of urns - note that neither card renders significantly more geometry than the other. ATI claims that supertiling works on virtually every D3D application. There are only two drawbacks - no OpenGL support and secondly it does not always give optimal performance and this is where the third rendering mode kicks in, Scissor Frame Rendering.

Object rendering will be split between two cards fairly evenly with tiling

Scissor Frame Rendering (ATI SFR) is similar to NVIDIA's Split Frame Rendering (NV SFR) but there are a couple of distinctions. ATI can split the frame either vertically or horizontally and two halves are not necessarily 50/50. From my interpretation of ATI's explanation, it seems that the ATI SFR implementation is not dynamic meaning that the split percentage is predetermined for each particular application. In contrast, NV SFR can only do a horizontal split but the load dynamically shifts to keep the workload between the two GPUs as even as possible.

Super AA

Because there is no real reason outside of benchmarking to play games with vsync off, ATI has introduced several new antialiasing modes to increase visual quality. Two of the new AA modes extends the most common form of antialiasing today, multisampling. Multisampling takes several points within a pixel and applies some blending to smooth out the edges. The more samples that are taken, the smoother the edge will look. Two of the Super AA modes introduced with CrossFire will double the number of sample taken and increases the multisampled AA modes from 4x and 6x to 8x and 12x.

Multisampling only versus Multi+Supersampling

Two additional AA modes exist and they take advantage of another technique called supersampling. Quite simply, the scene is rendered at a resolution that is higher than that of the display and the image is scaled back down. Supersampling is fillrate intensive and for that reason it has not been used for several generations of video cards. ATI's 10x and 14x Super AA modes does not simply do supersampling but use a blend of multisampling and super sampling to increase visual quality.

ATI's antialiasing patterns