Too slow speed for lammps gpu

[Juo-kim]

Hello,

I am running the mace LAMMPS gpu , but the computation speed seems unusually low. I am using an RTX 3090 GPU, and I will attach the log file and the file I used for compilation.

The structure I used contains only 12 atoms, so I am wondering if this is the expected performance.

For the MACE model, I downloaded and used the one named "2024-01-07-mace-128-L2_epoch-199.model-lammps" from MACE-MP.

And this is my input file
####input#####
units metal
boundary p p p
atom_style atomic
atom_modify map yes
newton on

read_data ./dat.lammps

pair_style mace no_domain_decomposition
#pair_coeff * * ./2023-12-03-mace-mp.model-lammps.pt Zr O
pair_coeff * * ./2024-01-07-mace-128-L2_epoch-199.model-lammps.pt Zr O
dump myDump all custom 10 xyz.lammpstrj id element x y z
dump_modify myDump sort id element Zr O

thermo_style custom step time cpu pe ke etotal temp press vol density
thermo 10

velocity all create 300.0 12345
fix myEnse all npt temp 300.0 300.0 0.1 aniso 1.0 1.0 1.0
timestep 0.002
run 10
################
lammps.log

lammps-mace compile.txt

[gabor1]

You need to run a few hundred maybe a thousand steps to get a good sense of speed, because the neighbour list calculation is not done on every step.

[gabor1]

I'm looking at the supplementary of the sevennet paper (https://pubs.acs.org/doi/10.1021/acs.jctc.4c00190) and they show that for a single GPU, sevennet is 1.5 times faster than mace (they don't say which mace, but it is likely the medium one)

[gabor1]

they also show that the multi-gpu implementation of mace is not efficient, which we know, and we are working on it.

[Juo-kim]

Yes, Considering that the speed may vary depending on the structure or environment of lammps, it seems to make sense.
Thank you for your attention!

[ilyes319]

@Juo-kim What kind of speed do you get with seven net on a single GPU in terms of ns/day? I would be very surprised is 7net is 5x faster than the L2 model for something like 200 atoms.

[liudfgoo]

Recently, I encountered the same problem. And I think it's related to the cutoff distance settings in the lammps. In the file uploaded by @Juo-kim, the neighbor/atoms and distance cutoff are below,

  update: every = 1 steps, delay = 0 steps, check = yes
  max neighbors/atom: 2000, page size: 100000
  master list distance cutoff = 14
  ghost atom cutoff = 14

Logically, the cutoff radius here should align with what was used in model training. And I want to know how to set the radius here. Looking forward to your help!