OSDev.org

Well what's mainly with 24 bits is that it needs a lot of small bits of time:
nonstandard address multiplications, misaligned memory accesses, non-dword writes, etc.
And you can't simply pack anything into multimedia registers with 24-bit colors... optimising 32-bits with MMX suddenly becomes trivial.

you can draw horizontal lines in 32-bit and 16/15-bit modes by using a single rep stosd/stosw
in 24-bit modes this doesn't work