i-PI + JAX/Torch models: efficiency

[Sander Vandenhaute]

Hi all,

How efficient is i-PI + PyTorch force evaluation compared to, say, directly running simulations with LAMMPS/OpenMM + PyTorch? Particle positions and box vectors need to be communicated at each step, which might take almost as much time as a force evaluation (which typically takes about 10 to 100 ms for a few hundred atoms)? 

Thanks,

Sander

[Michele Ceriotti]

The overhead from communication should not be too high. Of the order of 10ms if you set everything up nicely, and you use a Unix domain socket. This is mostly down to latency, so if you go to a larger system the communication overhead shouldn't grow substantially.

We are currently doing some profiling https://github.com/i-pi/i-pi/pull/277 so we should also be able to reduce a few bottlenecks and give better indications of how to set up a "fast" calculation. 

Michele

[Michele Ceriotti]

Just reviving this thread to point at the i-PI 3.0 preprint https://arxiv.org/abs/2405.15224 that discusses quantitatively the (negligible) overhead of using i-PI in precisely this scenario.