This is a FAQ — you don’t get any performance boost by “transpiling” from another language, because the whole performance benefit of Julia derives from the front-end, not the back-end.
SymEngine is reportedly hundreds of times faster than SymPy for many tasks, and Symbolics.jl is reportedly matching SymEngine already on many benchmarks (but seems easier to extend and integrate with other Julia projects than SymEngine’s C++ codebase).
To quote one of the authors of SymPy and SymEngine, Python is indeed too slow, so it is hard for SymPy to compete in terms of speed. The idea that a new implementation in a compiled language can do much better hardly seems outlandish.