By "make up" for it I mean you might get the same performance by moving more of your processing to the GPU and avoiding the CPU altogether. Obviously you can't do the same kind of close interaction between CPU and GPU like you could with shared, cache-coherent memory.

I can't wait until there's a PC that has that kind of memory architecture, though, if we ever get one.