TheRaven64's Journal: Is 3D Hardware still Hard?

First generation video cards were basically a blob of memory and a DAC. Then came simple 3D cards that were basically rasterisers. Then the cards started doing transform, clipping, and lighting calculation all in hardware. Then manufacturers started moving away from the whole fixed-function concept. Modern graphics cards are basically floating point vector stream processors.

While the Open Graphics Project is aiming to build a fairly simple fixed-function device, I wonder if this is the best approach. It seems that it would be simpler to start with something like the OpenRISC core and:

1. Remove the integer units (not needed).
2. Increase the floating point pipeline width to 128 or 256 bits, giving 4-way SIMD on every operation.
3. Add hardware support for trig operations.
4. Stamp a few dozen of these cores onto a die.
5. Do everything else in software.

You'd get a chip that is heavily optimised for doing graphical operations, but still a general purpose design, a lot like modern GPUs. The other benefit is that there are a lot more open source developers capable of writing software OpenGL implementations for this kind of device than there are able to make meaningful contributions to GPU design.

For bonus points, you could add a few stock OpenRISC cores, a memory controller, USB controller, etc to the design, and have a completely open source system-on-a-chip design.