3D is intrinsically parallel. We find parallelism at the level at which data is manipulated (4 component vector), but also with the instructions applied to that data. GPU manufacturers have understood that fact well and it's around this main idea (which constitutes the main difference between graphics and CPUs) that 3D chips are built.
Consequently, an increase in parallelism is the best way to increase GPU performances Compare Prices on ATI Radeon HD 3870 X2. Whether you're talking about geometrical units, pixel pipelines or steam processors, the history of the evolution of graphic card performances is largely related to the one telling the number of those units. The only limit is the physical size of the chip, the increase in which has slowly led to an explosion of production costs. In the end, optimization of processes is the only thing that allows us to increase the quota of transistors that a GPU architecture can hold (their size doesn't vary too much) and thus the number of computational units.
To outgrow this limit, the founding fathers of 3D introduced the idea of multiplying chips. Once again, this approach is largely beneficial in the GPU world, as tasks can be easily parallelized and assigned to a specific chip while still limiting the number of exchanges between the different GPUs, unlike what you might have seen with multi-CPU solutions. Thus the Infinite Reality from Silicon Graphics was conceived and meant to be largely configurable, the different parts of the graphical pipeline being implemented on independent cards: Geometry Engine, Raster Manager, Display Generator.
It could be set up using multiple configurations and it was even possible to add cards to increase workstation performances. The Rasterization diagram is complex and aims to intelligently balance the workload, either between the 80 image engines of each Raster Manager, or the different Raster Managers. To do so, it operates using tiles that vary in size depending on the number of Raster Manager.