This argument smells like bullshit, sorry, and it's totally not justified in the article. Giving just numbers without even explaining a potential causality points to benchmark setup failure more than anything.
Unless you are talking about statically linking your malloc implementation, your entire STL and all depending libraries AND then performing LTO on it, and even then I'd be surprised if there's any noticeable improvement over this.
Not to mention I don't know of _any_ real-life executable that does this...