I'm not aware of any Intel CPU on which it actually does invalidate
the entire TLB, although that doesn't mean there isn't one.
But you're looking at the architectural specification. In hardware it
may be (very) inconvenient to guarantee the absolute minimum number of
TLB entry purges, depending on how exactly the TLB gets indexed. So
they're just saying it may be more than the minimum*, and they've
specified the largest possible maximum.
Section 4.10 of volume #3A goes into this in some detail, especially
the requirements for when you have to do specific invalidations.
*For example, they list as specific examples that INVLPG may delete
TLB entries for the given page for different PCIDs, or that on
multi-threaded processors, TLB entries for one thread may be
invalidated by a different thread.