Plaz, you are not talking about AGP texturing (DME), you are talking about data transfer. The Voodoo3 for example uses AGP 2x but not AGP texturing. That is what I'm talking about.
OK. I think I finally understand your question!
:-)
When they say that the Voodoo3 is missing AGP texturing it means that the rendering pipeline can only apply textures that are loaded into the card's memory. It can't render textures straight out of system memory. Cards that support AGP texturing can.
Why is this significant? Because cards that support AGP texturing don't have to dump existing textures in their on board memory just to make room for that texture.
So say you have 14MB of free texture memory on your card, but your scene calls for rendering 18MB of textures. With AGP texturing, the extra 4 megs could just be rendered from system memory. Sure, there'd be a performance hit, but things would still keep chugging along OK. But on a Voodoo card, this could be really bad, if the example accessed the textures in a "perverse" enough order that caused the entire texture cache to be flushed every frame. Every texture would have to be loaded from system memory (at AGP2x speed) every time. Now the caching algorithms are somewhat intelligent, so this is a worst case scenario that probably wouldn't happen, but it could (for example, if you work for S3 and coded an example to do it :-) ).
There's no software trick to solve this. The V3 is wired such that it can only render textures from it's on board memory. It does transfer texture data from main memory to the card at AGP2x speeds, but it has to transfer them before it can draw them.
So AGP texturing is somewhat useful, but as the percentage of textures being rendered straight from system memory increases, performance decreases dramatically. I think this performance hit will be even more apparent on the next generation accelerators.
Plaz |
