Lattice Boltzmann method (LBM) is one of methods for aerodynamics simulations. Stability- and accuracy-enhanced schemes of LBM such as multiple-relaxation-time model or cumulant model have high arithmetic intensities of 10~100. They are suitable for high performance computing with GPU supercomputers. We apply an octree-based mesh refinement into LBM to assign high resolution mesh near complex shape bodies, and implement to a multi-GPU computation. We uses a domain decomposition based on rectangular space-filling curve. An aerodynamics simulation for racing bicycles with Reynolds number 1,000,000 is performed with 8 GPUs. The total mesh number of the simulation is 126,374,256 and the mesh resolution near the bicycle is 4 mm. It takes 2 days to simulate 10,000 time steps, that correspond to 0.11 sec in physical time. The results have good agreements with the experience on real bicycle racing.