[geeks] Itanium 32 bit performace.... hahahaha

[jdboyd at jdboyd.zill.net]

On Tue, Dec 09, 2003 at 11:19:13AM -0800, Francisco Javier Mesa-Martinez wrote:

> > Are you sure your information is still current on that?  I realize that
> > was the case as of the TNT2 era, and perhaps even to some extent the
> > Geforce 256 and Geforce2 era, but it would seem that the Geforce3s and
> > later change that.
> >
> > I wouldn't have expected the projects mentioned on GPGPU
> > (http://www.ibiblio.org/harrism/index.shtml) to bother if the geometry
> > was just being shoveled to the CPU to do the work.
> 
> Most of the transforms are done in the CPU, the GeForce line is geared
> towards raterization (and that is what they are good at).
> 
> There is a reason why 3Dlabs is still in business, or why you have to
> showel the big buck for a Quadro :).

You won't get any argument from me about a Quadro being worth the extra
bucks.   

However, while I don't have my precise numbers handy anymore, in some
crude testing of custom code that was transform bound* on a P2 system
running linux with a Geforce3 versus a TNT2, the code did far better
with the Geforce3.  For comparison purposes, it ran almost as well on
P2-350 with a Geforce3 as it did on a P4-1.6ghz machine with a
Geforce3 until I added collision detection (which ran much faster on the
P4, and made only minimal use of OpenGL to perform**).

Anyway, I'm not a major expert at OpenGL, and worse, I wrote the stuff
in the manner that SGI recommended for Indigo2/Impact machine more so
than was recommended by Nvidia.  That may have effected some things, but
I just didn't have time to deal with things like converting my models to
strips. 


* I determined that it was transform bound by noting that it ran no
  faster in a 320x240 window than it did in a 800x600 window.

** I used code from an OpenGL -> postscript demo to cache where every
   vertex was really located after the hierarchy of objects was
   transformed.  SGI seems happy with things like this, but some source
   suggested that it may be sub-optimal for AGP based machines, which I
   somewhat disregarded on the theory that the SGI way was the one true
   way.