2D Arrays texture or global for fermi cards?

Accelerated Computing CUDA CUDA Programming and Performance
#1

What is performing best in Fermi cards for 2D arrays?
Texture or Global memory?