Choosing Hardware for LAMMPS Simulation

Hello there! I hope everyone is doing well. I’m reasonably new to LAMMPS, and Molecular Dynamics is not my area of expertise, so I appreciate the patience.

I’m working on a research project that aims to show fluid-like behavior by simulating the interactions of many individual particles across extensive periods (on a molecular scale). In practice, this consists of simulating \mathcal{O}(10^7) argon atoms interacting via LJ in a periodic 2D domain for \mathcal{O}(10^6) time steps.

Fortunately, I have been granted access to a computer cluster at my institution, which, thanks to the current low demand, I may use with little or no cost.

That being said, I need help choosing which computing node would best fit my application. The nodes I’m considering are:

  • H100 Nodes with:
    • Dual Intel Xeon Platinum 8462Y+ CPUs @ 2.8 GHz (64 cores/node)
    • 2TB DDR5 DRAM
    • 4TB NVMe
    • 100 Gbps InfiniBand NICs
    • 8x Nvidia Tensor Core H100 80GB
  • A100 Nodes with:
    • Dual AMD Epyc 7513 CPUs @ 2.6 GHz (64 cores/node)
    • 512GB DDR4 DRAM
    • 2x Nvidia Tensor Core A100 40GB or 80GB GPUs
    • 1.6TB NVMe

Any advice on the matter will be greatly appreciated.

Then you are advised to find yourself a collaborator who has the required expertise so you can learn from them.

The H100 nodes obviously have more potential (8 GPUs instead of 2 and a newer generation). However, how well either of those work for your specific research project can only be reliably determined by making benchmarks with representative inputs.

Please note that the potential for GPU acceleration is somewhat reduced for 2d systems versus 3d systems since atoms have far fewer neighbors but the GPUs need many neighbors to have sufficient work units.

Since you are doing simulations of a simple LJ system, you should also consider other, more specialized MD code, which may give you even better performance than LAMMPS, like HOOMD-blue.

1 Like