It's to increase the effective memory bandwidth of the system. GPUs have very high memory bandwidth requirements, so the models with the high end GPU would be limited by the memory subsystem. Using the extra cache the amount of fetches from memory can be reduced, increasing the amount of memory bandwidth available.

That makes sense. Using it as a victim cache seemed kind of odd, but I think it makes more sense now. It seems like it's a still evolving technology. I wonder if there are corresponding performance counters.