As I said in another comment, probably the benchmark is done just using some hardware instruction that didn't exist on those models and gets compiled to several instructions (possibly by a very very old compiler, while we're at it) vs something handwritten in assembly for the purpose of one specific benchmark.
Does this mean it's 23x faster for normal workloads? Nah.
Apple when they were pumping clang were also claiming that binaries produced with clang were much faster than those made with gcc. This was because they used a 15 years old version of gcc that didn't have any vector instructions (because they didn't exist at the time) and benchmarking using some code that was solely doing vector stuff.
