Which is faster in CUDA: Constant Memory or Texture Memory?

I know that both is on off-chip DRAM and cached.

But which is faster in access speed? Or in what circumstances one is faster than the other?