Trying to verify the correctness of my optimizations by replacing all floating-point values with symbolic algebra objects. It turned out that getting a full symbolic trace of over 100x100x100 cells over 100 timesteps isn't exactly a great idea... Even a 20x20x20 grid is enough to fill 4 GB of RAM, so a 100x100x100 grid needs 500 GB of RAM...
@robryk@qoto.org Yes, it's also progressively getting slow in every timestep as well.