Andy got us some new numbers.  The tests he tried to run most recently
included GeForce 6200, Radeon 7000, a Matrox G550 (PCIe), and a Matrox
G450 (PCI).  The reason for the last two tests was to compare the
effect of PCI vs. a faster bus on two very similar GPUs.
Unfortunately the Matrox cards and/or drivers are too primitive to run
Quake, so that's out.

Here are his results.  The quoted figures are the 640x480 numbers,
while the new (slower) figures are for 1280x1024.

GeForce 6200 AGP 256MB:
quake3: 1346 frames, 7.0 seconds, 192.8 fps
glxgears: 1360.8 fps

For 1280x1024, the GeForce drops down to:
1346 frames, 23.7 seconds, 56.9 fps

Sapphire Atlantis 7000 PCI 64MB DDR:
quake3: 1346 frames, 43.4 seconds, 31 fps
glxgears: 343.584 fps

For 1280x1024, the 7000 drops down to:
1346 frames, 137.7 seconds, 9.8 fps


I'm going to offer up an assumption for us.  I'm going to assume that
Quake tries to show us the same image, regardless of the resolution.
To do that, it would paint the same scene, using the same number of
triangles.  So when you go to a higher resolution, the same number of
commands is sent to the GPU, but each command results in drawing more
pixels.  This assumption is probably wrong, but it's probably close
enough to the truth that we can use it as a rule of thumb.  Going
based on this math, 1280x1024 has 4.27 times as many pixels as
640x480, so the higher res should have a framerate that is 1/4.27 of
the lower res.

The GeForce (on AGP) dropped in framerate by a factor of 3.3, which is
in the right ballpark from the prediction.

The Radeon (on PCI) dropped by a factor of 3.1, which is still in the
right ballbark, unfortunately.  This means that either PCI is fast
enough to keep the Radeon busy, even at 640x480, or my hypothesis is
wrong.

Anyhow, given this incredibly limited amount of information, I'm going
to throw out an initial projection that OGD1, with the OGA1 GPU, will
have a framerate of somewhere between 20 and 30 FPS on Quake III at
1280x1024.  The 20 comes from the fact that the low-end projection for
OGA performance has twice the pixel throughput as the Radeon.  The 30
comes from the fact that OGD1 has 3 times the available memory
bandwidth (a pixel often requires more than one memory access).

If we can prove that PCI is the bottleneck here, I'll have to reduce
those numbers.
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to