Why cache working set per multiprocessor for texture memory has a variable size?

I saw it here Programming Guide :: CUDA Toolkit Documentation