Performance comparison of LAMMPS MLIAP and Symmetrix across NVIDIA GPUs

LAMMPS LAMMPS General Discussion
, , , ,
1

I am running molecular dynamics simulations using MACE-MATPES-R2SCAN-0 and comparing the standard LAMMPS MLIAP interface with the Symmetrix MACE evaluator.

The representative systems contain 960 atoms, use a 0.5 fs timestep, and follow a 3.5 ns heating/cooling protocol. Each calculation uses one GPU, one MPI rank, one OpenMP thread, and Kokkos:

-k on g 1 -sf kk -pk kokkos newton on neigh half

The direct calculations use:

pair_style mliap unified

The accelerated calculations use:

pair_style symmetrix/mace

Performance measured from the LAMMPS logs:

  • H100 with Symmetrix: 2.442 ns/day (2.404–2.465)
  • A100 SXM4 40GB with Symmetrix: 1.353 ns/day (1.324–1.403)
  • L40S with Symmetrix: 0.783 ns/day (0.774–0.805)
  • RTX PRO 6000 Blackwell with direct MLIAP: 0.527 ns/day (0.507–0.546)
  • L40S with direct MLIAP: 0.368 ns/day (0.359–0.384)
  • GTX 1080 Ti with Symmetrix: 0.213 ns/day from one production run

Has anyone systematically compared the performance of direct MLIAP and Symmetrix for MACE foundation models across these GPUs? Any suggestions for improvement?

2

Has anyone systematically compared the performance of direct MLIAP and Symmetrix for MACE foundation models across these GPUs?

Would make a great blog post or github repo if you made a quick chart and shared your code

3

I cannot say I have, but I can add a new piece to the discussion. Fixing the type of atomic system (HEA) and the type of GPUs (GH200) I got the numbers I am showing here: ML-IAP inference performs substantially worse than Symmetrix on ARM64 and CUDA13 · Issue #1708 · ACEsuit/mace · GitHub

So the only thing I can say is that ML-IAP is generally slower than Symmetrix (marginally faster when running “large” systems on a single GPU), and that for some reason it does not really scale (neither in the strong sense, nor in the weak sense).

Any input from someone who actually know what is going on is more than welcome!