Multi-GPU Solutions in the Past, Present and Future

Again, this multi-GPU approach must be a temporary thing until ATI and Nvidia are prepared to launch their next-generation GPUs. We produced arguments in favor of this point of view in our Radeon HD 3870 X2 review. In brief, the developer company ensures the freedom of price and market maneuvering for itself. On one hand, the company can utilize defective cores by disabling some subunits in inexpensive graphics cards. On the other hand, as the tech process improves, the frequency potential and, accordingly, performance grow up. This helps extend the product range to better serve the end-user. A transition to homogeneous multi-chip solutions is no good for both parties: for the manufacturer because such graphics cards are expensive to manufacture and support, and for the buyer because there would be few solutions to choose from, and they would be more expensive in comparison with single-chip solutions with similar characteristics.

Well, the single-core concept has a serious drawback, too. The manufacturers cannot increase the performance of homogeneous graphics cores infinitely because it makes the chip too complex and too hungry for power. We’ve talked about the downsides of homogeneous multi-chip solutions a lot in our reviews. Perhaps the future belongs to heterogeneous Multi-Chip Modules. The concept isn’t new, actually. Texture processors and rasterization units were designed as individual chips back at the times of the 3dfx Voodoo and Voodoo 2. This approach was also employed in professional 3D accelerators from 3Dlabs in which, for example, a geometrical coprocessor could be an individual chip.

On a new level, this could be a set of dies of varying functionality combined in a single package and connected with internal high-speed interfaces. For example, one die contains a command processor, ROPs, memory and system bus controllers, and DVI/HDMI/DisplayPort controllers, and the other die contains TMUs and universal execution units. Another possibility, the basic die may only contain a command processor and controllers of I/O interfaces while everything else is implemented as additional dies. This would provide even more flexibility of configuring such a solution. Performance can thus be increased easily by installing additional execution units into the chip package. This approach is less profitable than the single-die architecture, but looks preferable to the homogeneous multi-GPU concept irrespective of implementation (MCMs or several chips with different functionality). The main advantage of a heterogeneous MCM/Multi-GPU product over homogeneous MCM and classic multi-GPU ones is that it always delivers maximum performance and does not depend on software optimizations. No wonder as it is actually an ordinary graphics processor whose functional units are distributed among several physical chips. This approach seems to be the most progressive, but it’s too yearly yet to predict its future.

In this review we’ll describe the Nvidia GeForce 9800 GX2 graphics card and 4-way SLI configurations, but we’ll also cover the current state of multi-GPU technologies to see what benefits they can bring to the gamer. We’ll discuss classic dual-chip solutions as well as graphics subsystems with multiple GPUs, including quad-processor Nvidia Quad SLI and ATI CrossFireX configurations.