The paper provides a brief comparative overview of the potential and limitations of the RANS and LES approaches to the computation of free and forced conveclive heat transfer. Major deficiencies and possible improvements of RANS models are discussed based on a priori validation against DNS data and computations with various models. A genuine niche for LES is illustrated by successful prediction of low-Re-number conjugate heal transfer over internal wall-mounted protrusions imitating electronic components. Limitations to LES at high Re and Ra number heat transfer are revisited and possible hybrid RANS/LES routes are discussed. The potential of Transient RANS for high Ra numbers thermal convection with pseudo-deterministic structures are illustrated in some generic and real-life examples.