The Pixel-Plane 5 is an ongoing project at the Department of Computer Science at the University of North Carolina at Chapel Hill by Dr. Henry Fuchs' group. This project has been carried out for the past 10 years. The machine was designed to be a super fast graphics machine, capable of real-time manipulation of polygons and other simple geometric primitives. One of its goal is to make Virtual Reality possible.
The heterogeneous graphics architecture is a composition of a fast RISC based distributed memory MIMD machine, the graphics processors(GPs), and thousands of SIMD elements, the renderers, connected in a ring network. Figure 7 illustrates the architecture.
Figure 7: Pixel-Plane 5 architecture.
High level rendering functions, such as keeping the data structures for objects, applying transformation matrices, clipping, modeling light, and so on is done by the GPs. When the GP is done with its set of calculations, it sends the graphical primitives to the renderers that scanconvert the primitives and send the final image to the frame-buffer.
The renderers act as a smart frame buffer. They are grouped in a
128x128 grid. They have the ability to automatically scanconvert
primitives by using a device called the Quadratic Expression Evaluator
(QEE), that is used to feed the renderers with their data, this device
can evaluate the function 
at each
position. They can also perform z-buffer and compositing operations.
After the renderers finish with their computation, the data is taken
to a frame-buffer. If enough patches are available to cover the screen
resolution (usually 1280x1024), the whole image can be scanconverted
at once, otherwise scanconvertion can be done in patches.
There is a fully functional prototype of the Pixel-Plane 5 available at the University of North Carolina, it uses 40Mhz i860 as the graphics processors and custom build SIMD processors for renderers. A new machine called Pixel-Flow is currently under construction, its architecture and functionality greatly resembles the one just described.