3d graphics multiprocessor

[Jason Watkins]

Apologies if this is considered off topic, there are few communities around
that deal with gfx hardware design.

I've been wondering for a while why someone doesn't develop a graphics chip
based on a multiprocessor design. Take an appropriate memory architecture, a
decent performance IEEE single precision core, and stir in the rest of the
DAC, bus control, etc you'd need.

The concept being that for a small R&D effort, spending time on an instanced
ALU and implimenting algorithms in software where they are easily revised
without hardware costs would be a winning strategy.

I've found some work on this from the stream multiprocessor crowd, however,
their memory architecture handles random access to small working sets
poorly, something that is *very* nessisary for good fill rate from texture
mapping. I don't think there's any real reason a stream like architecture
could be combined with appropriate caches.

So I was wondering if any of you had thought along these lines, knew of work
along these lines, or immediately know why it is an approach doomed to
failure.

jason watkins