Your Favourite method of Driver distribution of load over multiple GPUs

We have here good and old:
http://en.wikipedia.org/wiki/Tiled_rendering
&
http://en.wikipedia.org/wiki/Scan-Line_Interleave


Then we have:
http://en.wikipedia.org/wiki/Scalable_Link_Interface
&
http://en.wikipedia.org/wiki/ATI_Multi_Rendering

When I look at 1920x1080 screen it has 2073600 pixels.

• Tile Based uses blocks of 16x16 = 256 pixels, which results in 8100 blocks which can be distributed pretty evenly over multiple GPUs.
  Scan-Line_Interleave splits by lines which results in pretty good 1080 blocks
  Scalable_Link_Interface&ATI_Multi_Rendering are those who just butcher areas somehow based on number of GPUs (2~4)

Evenly distributed workload of those modern technologies very much depends on how flexible driver is in decision making (what to give to each GPU).

To me, games which needs SLI/CF are those shader heavy beasts, so little scene work (geometry, z, ...) which is duplicated is not that bad cost to get even and Full workload for any number of GPUs.

What is it to you?

 

Spot on. Tiling is awesome if done well. We'd be able to use more than 4 GPUs too (without worrying about mixing AFR and SFR); essential for some scenarios like triple 2560x1600 monitors @ 120Hz maxed out

 

It is a little more complicated than that when you have things such as dynamic lighting where you have to know the effect from different areas' lighting for the area being rendered.

 

Complicated, sure, 10 years ago there were already grid rendering applications.
It's so complicated, that nV has farm capable to render in real time fotorealistic images with full ray-tracing. And it distributes work over few dozens of GPUs.

It's not easy topic, but software solutions were perfected many years ago. Just implementation in nV/AMD consumer drivers lags a bit.

And to the lightning, it's not like you render part of scene which is illuminated and then takes approximate intensity from that and scatter it around... It would be total waste of rendering cycles.

Complete scene is used anyway in all methods with exception in AFR where geometry is processed only once per rendered image. But I consider that method as weakest personally.