Comment by fogleman

It's explained in the README. The SDF is evaluated on whole batches of points (32x32x32 tiles by default) using numpy. (And it's multithreaded.)