You are talking about different things.
You have the modelling language (YALMIP vs LMILAB language), and in my opinion the LMILAB modelling is absolutely horrible and obsolete. It is extremely cumbersome to model anything non-trivial. The whole reason YALMIP was developed some 20 years ago was the agony of doing anything advanced efficiently in LMILAB, and it is just as bad today.
You then have the solver (interfaced via YALMIP such as mosek, sedumi, etc, vs the solver in LMILAB). The solver in LMILAB is several decades old, and pretty inefficient on general problems. It is efficient on standard Lyapunov-ish LMI control problems, if modelled using the specific modelling constructs in LMILAB, but the model has to be really large and simple for this to be any kind of motivation for struggling with the modelling language LMILAB. The solver dos not have any constructs to exploit SOCP cones either, so it is a very limited solver compared to solvers sucas mosek,sedumi, sdpt3 when you start doing anything but vanilla control LMIs.