@akohlmey, this is the figure from Grime and Voth (2014) that got me very interested in atom_style template :
plus it doesnt make sense to store duplicate topology information millions or billions of times. im working with large systems (that’s why im with lammps in the first place because gromacs has the 32 bit barrier)
my first minimal test case has 2.3M molecules of only 5 types. i plan on scaling way past that so any algorithmic savings in memory are worth my effort, including porting to atom_style template/kk and debugging/fixing anything that needs it along the way.
but not GPUs, that’s the limiting resource to scaling for me at the moment
maybe… but even LAMMPS scales better at O(1) than O(n^2) !