ATI’s SuperTiling ?Best Load Balancing Ever?

ATI CrossFire offers four new rendering modes with different peculiarities:

SuperTiling ?image is split into 32x32 pixel squares, half of the squares are rendered by one graphics card, another half is rendered by another GPU. ATI claims that such chess-board like approach provides superior load balancing, which seems to be correct, as it does not require any intelligent driver to balance the load, yet, given that tiles are relatively small, parts of the image with nearly equal complexity of rendering are distributed equally to different graphics cards. SuperTiling works in Direct3D applications only and requires both graphics cards to store identical frame buffers and process all the geometry for the frame.

SuperTiling example. Click to enlarge

Scissor ?image is split into two parts, driver dynamically analyzes the frame and balances the load between two graphics cards. The scissor is supported by both Direct3D and OpenGL applications and works just like NVIDIA’s Symmetric Multi Rendering technology. The approach also does not provide any geometry performance advantages as both cards process vertexes for the whole frame, according to ATI.

Alternate Frame Rendering ?already known ATI’s technology that was used with the company’s Fury MAXX graphics cards. In the AFR process, one chip renders even frames while the other chip renders odd frames. Each chip processes triangle setup for its own frame without waiting for the other chip, which makes AFR very efficient in terms of performance scaling. AFR is supported in OpenGL and Direct3D.

Note that neither of the approaches actually allows geometry scaling, which limits the number of advantages multi-GPU technology can provide to a couple of general cases:

• Pixel shader performance scales well with addition of GPUs;
• Fillrate limitations are avoided thanks to multi-GPU.

Generally speaking, geometry performance is not something needed tremendously for modern games. The only benchmark that can benefit from additional vertex shader power is probably 3DMark05. Furthermore, current ATI hardware has certain limitations in pre-vertex shader buffer (which is currently a 256-bit line), which means that even with a single visual processing unit geometry performance may not be limited by vertex processors themselves.

Super AA ?RADEON graphics cards work independently rendering image using different FSAA masks. CrossFire Compositing Engine then blends the two images for enhanced image quality with real-time performance. Super AA mode is enabled through the control panel.

With CATALYST A.I. enabled, the preferred rendering mode is selected for targeted applications automatically. For applications that are not identified in CATALYST A.I., or when CATALYST A.I. is disabled, default multi-GPU rendering modes are offered.

By default either SuperTiling or Scissor modes are applied. Alternate Frame Rendering mode is used for applications identified in CATALYST A.I. (when enabled). When CATALYST A.I. is disabled, graphics processors with 16 pixel processors running Direct3D applications are accelerated by SuperTiling mode, whereas other hardware configurations (e.g., 12-pipe GPUs) are accelerated by Scissor mode.