As opposed to Nvidia in 2006, the former ATI Technologies doesn’t have such an opportunity in 2008. The ATI Radeon HD 3870 X2 must work ideally to be successful on the market whatever its price. The new card is schematically similar to the Nvidia GeForce 7950 GX2 but uses CrossFire rather than SLI technology:
The common thing between the new card from ATI and Nvidia’s solution is the use of a PCI Express switch that allows for a direct communication between the GPUs. The switch installed on the card complies with the PCI Express 1.1 specs, so the data-transfer speed is 4GB/s in each direction, although the RV670 supports PCI Express 2.0 and a twice higher communication speed. On the other hand, the capabilities provided by the PCI Express 2.0 standard are yet not called for by real applications, providing no performance boost, just like it was with AGP 8x.
One of the CrossFire channels is used for communication between the Master and Slave GPUs. The other is attached to the standard connector for the graphics card to support CrossFireX Quad configurations. Ironically, such a configuration is going to be much alike to the notorious Nvidia Quad SLI platform. Well, it may have a better fate because ATI’s multi-GPU support is implemented better than Nvidia’s, at least at the current moment.
The two GPUs of the ATI Radeon HD 3870 X2 card operate in the same way as two Radeon HD 3870 cards united in a CrossFire tandem.
The ATI Radeon X800 and X1000 used to support different multi-GPU modes, each with its own highs and lows:
- Scissor mode was the same as Nvidia SLI’s SFR mode. Each frame is split in two, and each part is rendered by an individual GPU. The size of the two parts may be varied dynamically by the driver to load the two GPUs in the same measure. Geometrical information is copied for each GPU, which has always been the main drawback of this mode.
- Alternate Frame Rendering (AFR) is a multi-GPU mode in which one GPU renders each odd-numbered frame and the other GPU renders each even-numbered frame. That’s the most efficient mode in terms of scalability because each GPU processes the geometry of its own frame and there is no doubling of the data. This mode is compatible with both Direct3D and OpenGL applications but used to require an application-specific profile in the ATI Catalyst driver to work.
- SuperTiling mode has no counterpart in SLI technology. In this mode the frame is tessellated into 32x32-pixel tiles. One half of the tiles are processed by the Master GPU, the other half, by the Slave. This is a fully hardware mode that does not depend on the driver. The geometrical data are duplicated for each GPU. This mode was compatible with GPUs that had 16 pixel processors and only worked in Direct3D applications.
Currently, the ATI Radeon HD 2000/3000 series do not support multi-GPU modes other than Alternate Frame Rendering although the ATI Radeon X1000 series used to work in Scissor and SuperTiling modes by default and enabled AFR only if the game profile was available in the driver.
Besides, Super AA modes are supported still. As opposed to the other CrossFire modes, these are intended to improve not performance but image quality by introducing additional FSAA modes (8x, 10x, 12x and 14x).
When you enable 8x or 12x FSAA, each GPU in the CrossFire pair works in 4x or 6x FSAA mode, respectively, but the samples are taken from different parts of the image and the frames are then combined into one to achieve FSAA equivalent to 8x or 12x. The 10x and 14x FSAA modes are a mix of traditional 4x/6x multisampling with additional antialiasing equivalent to 2x super-sampling. The number of texture samples in Super AA mode is doubled, which allows using 32x anisotropic filtering. Super AA modes are available in Direct3D as well as OpenGL applications.
We don’t know the reason for the changes in the CrossFire modes and if the SuperTiling and Scissor modes are going to be supported in the future but CrossFire technology may now prove to be even more dependent on the driver. There can also be the notorious AFR lags when one GPU outputs a frame that does not agree with the gamer’s actions one or two frames before.
Theoretically, AFR mode is the best one in terms of scalability because each GPU processes one frame, which eliminates the need to duplicate data, for example geometrical data. There is one thing that makes the whole picture not so bright, though. Besides the mentioned lags, data must be exchanged actively between the graphics core and the driver when the rendering of one frame depends on the results of the rendering of the previous frame. This is usually avoided by introducing an appropriate profile into the driver so that AFR could be effective. But if there is no such profile, the GPU array, or the ATI Radeon HD 3870 X2, switches into single-GPU mode with the appropriate performance loss. That’s not what you want from a new and expensive graphics card, of course.
So, the success of the ATI Radeon HD 3870 X2, and any other multi-GPU card such as Nvidia GeForce 7950 GX2, Nvidia GeForce 9800 GX2 and S3 Graphics ChromeMulti, depends largely on the programmers who are writing AFR profiles for as many games as possible.
Our tests will show how AFR is supported currently for the new graphics card. The next section is about the hardware of the Radeon HD 3870 X2.