Particle Mesh Ewald splits electrostatics into a short-range part, which decomposes cleanly across ranks, and a long-range part computed on a grid with FFTs. The FFT is global — every rank needs data from every other — so for years it stayed on a single GPU while the rest of the step spread out. Add more cards and that one grew into the whole runtime.

Recent GROMACS releases decompose PME across GPUs as well, with distributed FFTs and direct card-to-card communication instead of a round trip through host memory. On large systems across many nodes that is the difference between scaling stopping at a handful of GPUs and continuing well past it.

A protein in water — a hundred thousand atoms, give or take — saturates one good GPU long before it needs a second. The honest advice is to fill one card first: check that mdrun reports the GPU busy rather than waiting on the CPU, and only then think about more hardware.