To assess the runtime performance gain from incomplete solves, I ran a 256 x 256 pixel LES cloud field with dx=25m for 100 timesteps at dt=10 sec.

The simulations use 512 cores, i.e. 16 nodes with 32 processes each.

Compared are shortwave and longwave solve times for:

• 2str: 1D Twostream
• 3_10 TenStream with default ILU preconditioner
• 3_10 TenStream with explicit solver and max_it 1, full solve on spinup, etc.
• 2str with rrtmg spectral integration instead of repwvl

Outlook: if the new spectral integration scheme and the incomplete solves hold up (currently investigating) we have the potential to get 3D radiative transfer as cheap as currently employed 1D schemes. Stay tuned!