I would like to chime in my experience. For my applications it turned out that Julia’s non-sparse algorithms are very good (especially tricks with @view, mul!(), div!(), etc.) that I couldn’t beat it with any sparse implementations.
Remember, that apart from doing the sparse calculations your code probably has also other things going on (potentially a lot of them) that require also converting to and from for those calculations, which also takes time.
Here is a lengthy topic of mine which, deja vu, I was again navigated through by the generous help and time of @gdalle.