Chez Scheme, which implements R6RS, is the fastest. See https://ecraven.github.io/r7rs-benchmarks/

I would say the reason it beats Chicken and Gambit is that it has a Scheme-specific compiler to machine code that has been worked on for a long time with performance in mind.

Dybvig's talk on Macro Writers' Bill of Rights explains some of the optimizations Chez does: https://www.youtube.com/watch?v=LIEX3tUliHw (Why macro writers? Because the compiler will handle superfluous generated code so macros doesn't have to)

See also The Development of Chez Scheme for the history of it: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.72....