In this contribution, a single-node multi-GPU thermal lattice Boltzmann solver is presented. We implement a simplified version of the hybrid model developed by Lallemand and Luo in 2003, which combines multiple-relaxation-time lattice Boltzmann for the fluid flow with a finite-difference method for temperature. The program is based on the TheLMA framework which was developed for that purpose. The chosen implementation and optimisation strategies are described, both for inter-GPU communication and for coupling with the thermal component of the model. Validation and performance results are provided as well.