[ ... PXG and i860 ... ]
> It seems to be a good choice of processing power for the job at hand,
> I'd really be interested in knowing how the PX/PXG's work.

i860 itself had enourmous processing power.

PXG i860 is driven by passing 'command packet' in reserved RAM area in
which both CPU and i860 share.  The design and approach was pretty
common when CPU was slow and acceleration (3D geometry transformation
bozos) was implemented with a dedicated external processing power. 

The issue is, the true bottle-neck resides in the bandwidth of _data_
transfer between main memory and framebuffer.  PXG framebuffer is not
accessible by CPU, and the 'pipe' to send _data_ from main memory to
framebuffer is narrow.  The negative impact simply becomes evident
when the bunch of large/small mix of 2D images are about to be shown
on screen.  That's the reason why Ultrix FAQ recommends not to use
Xserver on PXG.  See also a paper how Digital guys managed to produce
useful Xserver with PXG.  (it should be available in Digital ftp site
under TriAdd/ directory)

Tohru Nishimura