It's probably because of the deprecation of 3dnow and/or a slower internal method emulating "obsolete" instructions like MMX/SSE.
Eg Iirc mmx and sse allowed 4 numbers to be operated on in parallel with olny one command to do so.
Eg "add 1 & 2, 3&4, 5&6, 7&8" would result in 3,7,11,15
If that instruction lost its optimised path, it could still exist as a valid instruction but take 4x longer (in clock cycles). Which hardly anyone would notice given the last software written for it was running on hardware at least 1/4 of the speed.
"3DMark 99 MAX is the first benchmark ever that has been optimized to fully support both Intel Pentium III SIMD and AMD 3DNow! instructions at the engine level."