On Wed, 3 Apr 2002, Raystonn wrote:

> A GeForce 3 produces about 88 million triangles/second maximum.  Now take a
> 3GHz Pentium 4, just a few months down the road.

So you are comparing a 2 generations old GPU with a CPU that won't exist
until the GPU's have cycled through at least one more generation.

OK - that's fair....NOT!

You should be comparing a 2GHz PIV with a GeForce 4 Ti 4600.

>  It has available 3 billion
> clocks each second.  This allows about 34 clocks per triangle.  With a
> properly optimized algorithm, using SSE and SSE2 to work on a tile at a
> time, this can be achieved.

Show me!

I don't believe you can rotate/translate/perspect/backface-cull/
light (with 8 light sources please)/clip (include two user clip planes
and six system planes)/texgen, and render several dozen multi-textured,
fogged, zbuffered and antialiased pixels - all in 34 clocks.

I want to see the program that can do this - render a complete
polygon from 'glBegin...glEnd' to pixels in the frame buffer
in 34 Pentium clock ticks.

Now add in multitexture and vertex/pixel-shader programmability
and the full richness of OpenGL per-pixel functionality.

>  What the main processor lacks in physical
> resources it can make up with a higher clockspeed.  I believe the GeForce
> 4Ti is at 300MHz now, is it not?

Dunno - but it has internal pipelining so that many operations happen
in parallel.  In a GPU you can be calculating perspective for one
vertex while clipping the previous one and lighting the one before that.
You only have to have a 10 stage pipeline and you have an effective
3GHz operation rate - I'd bet that modern GPU's have at least 100
pipeline stages.

> Fillrate is still an issue, but will
> become less of one as time goes by.  Memory bandwidth is ever increasing.

See my previous explanation of why this is so far from the truth.
In summary: Fillrate is becoming *MORE* of an issue - and memory
bandwidth is increasing *painfully* slowly compared to the rest of
the system.

> Use a scene-capture algorithm and you reduce the required fillrate
> dramatically.

But if those algorithms are practical for modern application
software (HINT: They aren't) then GPU's will evolve to do that
too.

I've built machines with "scene-capture" (look here:

  http://www.sjbaker.org/steve/personal/stargazer.html

...that's a screen capture machine).

In any case, for 'modern' polygon sizes, the best of these
technologies only improve fill rates by a factor of between
2 and 4.  Software has to go at least 100 times faster to catch
hardware fill rates.

----
Steve Baker                      (817)619-2657 (Vox/Vox-Mail)
L3Com/Link Simulation & Training (817)619-2466 (Fax)
Work: [EMAIL PROTECTED]           http://www.link.com
Home: [EMAIL PROTECTED]       http://www.sjbaker.org


_______________________________________________
Dri-devel mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/dri-devel

Reply via email to