Please provide a complete example I can use to reproduce the problem. Otherwise, I don't really know what you're doing or why it isn't working!

Can you provide a complete, working example that demonstrates the problem? The code you posted before doesn't correspond to what you described since it doesn't use a real integrator, doesn't run any dynamics, and never retrieves the positions.

simulation.context.setPositions(positions*1e+10*angstrom)

Is that really what you mean? I don't know what the values in positions are, but unless they're all tiny values on the order of 1e-10, you're setting the positions to huge values.

The correct command to test it is
Are you looking at an old version of the documentation?

When I load your PDB file, I get a lot of warnings like /Users/peastman/miniconda3/envs/openmm/lib/python3.9/site-packages/openmm/app/internal/pdbstructure.py:537: UserWarning: WARNING: duplicate atom (HETATM 6441 C1 HEXAB9999 4.711 32.046 48.163 1.00 10.00 C , HETATM 6421 C1 HEXAB9999 10.934 22.748...

In CHARMM, patches are generally applied at the modelling stage. If you used a patch while generating your PSF file, it should contain everything it needs to be read correctly.

Every atom in a residue is required to have a unique name. Because your atoms are all named either "H" or "C", it assumes they are alternate locations for the same two atoms rather than different atoms.

No, it does not use tensor cores. We've looked into it, but there isn't really anything they would be useful for. They have a couple of problems. 1. Really all they do is matrix multiplication. They aren't useful for anything that doesn't look like a matrix multiplication. 2. They don't support 32 b...

What do you expect to happen in that time? Is it a stable molecule or a very flexible one? 1 million steps is 4 ns. How much conformational change do you expect it to undergo in that time?

You're better off just using a single GPU. For one thing, 100,000 atoms isn't really that large. It probably doesn't provide enough work to saturate multiple 4090s. For another thing, the performance of multi-GPU simulations depends strongly on the speed of communication between GPUs. All consumer G...