performance issues

[Mirror]

I'm bringing back a performance issue i'm having, maybe someone would be kind enough to enlighten me.

i have a p4@2.8ghz, 768 mb ram, nvidia 6600 ( agp 8x as it says here : http://www.nvidia.com/page/geforce_6600.html ) but my motherboard bus is agp 4x. with just loading meshes - JUST meshes in wireframe mode and no other operation - i'm getting very low fps. for example for a sphere of 40k polys i'm getting 17fps, that means 680.000 polys / second which is very low.
with some non-sphere meshes, like for example plain grid of 80k polys i get about 14 fps which means 1.120.000 polys /second.

in some previous topic, hybrid had said that he also has an nvidia 6600 and with a cube of 50k polys he gets about 300 fps that means 50.000*300=15million polys /second which is a tremendous difference from mine. and even if half agp bus speed is a factor, how is it possible that such a big difference exists ? from 15m down to 1m.

Could it be that my low agp bus which is not 8x but 4x is a factor that contributes to this result ? i have the latest nvidia drivers, windows XP SP2. I repeat, im just rendering the mesh in wireframe mode, nothing else.

also i have a maybe stupid question.
http://www.nvidia.com/page/geforce_6600.html
in this link again it says that nvidia 6600 gt can handle 375 million vertices / second. that is 125million polys/second which is WAY far bigger than even hybrid's 15million/second. Is there something that i miscalculate or it is just that irrlicht does not use the lowest level and thus fastest d3d/opengl functions for rendering ?

[hybrid]

I did this number with an artificial, manually tweaked setup using the VBOs from SVN. I did create a custom scene node to avoid the possible render state changes and VBO re-enablings (because we don't have instancing, yet), which gave me around those numbers. And yes, that was a pretty good number

[BlindSide]

Basically, if you use VBOs, AGP bus speed wont effect it one bit, because the information will be stored directly on the video card's ram.

So to sum it up, use the latest SVN trunk (Has support for OGL and DX VBOS) and set all the meshbuffers of your meshes to use EHM_STATIC, E_HARDWARE_MAPPING.

EDIT: One more thing, wireframe mode is very very slow, so it will run faster normally.

Cheers

[sio2]

Consumer hardware may draw lines as degenerate triangles. Wireframe mode as a means of testing performance is not going to yield good results.

"375 million vertices / second" is probably the theoretical max fetch rate from the vertex cache.

Also, don't forget that the vertex cache only comes into play if you use indexed primitives!

Maximum throughtput probably requires degenerate non-textured, non-depth-buffered triangles.