That's because the card can't allocate anymore threads in the dedicated vram buffer without losing performance or tossing out errors.
The amount of shaders the 780 has makes it necessary to use a high -L (lookup gap) value to reduce the memory footprint of each gpu thread being spawned to get as many of them as possible to saturate the shaders, otherwise you'd be limited to smaller launch configs, ergo, lower thread count (think lower intensity on AMD gear) before getting HW (or cpu does not validate result kind of errors and crashes on cudaminer).
Lookup gap makes it possible to get to 7.5kh/s while using 2.3gb of vram, fine, but it also introduces a huge performance penalty compared to a lower -L value.
It would boost performance tremendously for instance, allowing us to use bigger launch configs, going from something like 3760 gpu threads at -L 6 to 7936. So yeah, perf would increase radically