Vertex Shader GPU Performance vs. CPU Performance

[Charlie Chang]

I have expected 10x difference. However, in our testing both had the same performance. Maybe thoushad vertices can show difference.

 

[Inane_Dork]

If you got the same performance, maybe you're bottlenecked somewhere else?

 

[Cryect]

Well the issue is if you are using the CPU for T&L then you can't exactly be using the CPU for lots of other things (sound decompression, AI, physics, etc)

Also I doubt vertex shaders are as possible as they could be. They really don't need to bother making them more powerful for PC's since games on the PC are pixel heavy and not poly heavy. And of course this is a chicken and egg type problem since hardware vendors won't waste silcon space on more powerful vertex shaders unless there are games for them and games won't bother with tons of polys if the framerate slows down and they can't get the full pixel processing power out.

Thats what will make unified shaders nice, instead of having to a specific balance of polygons to pixels instead the game developer can choose do they want tons of polys, tons of pixels or somewhere inbetween (current video cards lean towards the tons of pixels while the PS2 leans towards tons of polys).

 

[swaaye]

Well I'd say that ATI having all 6 vertex shaders even on X700 shows a sign that they are pretty serious about vertex thruput. NV is taking a different stance right now though.

 

[Humus]

I have expected 10x difference. However, in our testing both had the same performance. Maybe thoushad vertices can show difference.

A thousand vertices is too few to really show any difference. The bottleneck will likely be on fillrate. Try a few hundred thousands.

 

[Ostsol]

Better yet: if vertex performance is all you care about testing, render everything outside the view frustum.

 

[darkblu]

Better yet: if vertex performance is all you care about testing, render everything outside the view frustum.

alternatively, if you want to test both the vertex shaders and the tri-setup keep things in the view frustum but set your zbuffer to 0 and zfunc to less <- early z should like that.

 

[Charlie Chang]

Well the issue is if you are using the CPU for T&L then you can't exactly be using the CPU for lots of other things (sound decompression, AI, physics, etc)
...

Thats what will make unified shaders nice, instead of having to a specific balance of polygons to pixels instead the game developer can choose do they want tons of polys, tons of pixels or somewhere inbetween (current video cards lean towards the tons of pixels while the PS2 leans towards tons of polys).

Good comments. SM4.0 may be in the unified shaders. Leaning towards less polys may be favior in the future. More and more real works are done on the texture mapping. The bandwidth may an important benchmark in the future.

Guess most silicons of vertex shaders have been used in low efficiency.