LAMMPS bulk benchmarks

We run molecular dynamics simulations of bulks using LAMMPS to benchmark the simulation times on various CPUs and GPUs. Also, the computational performance of different force-fields might be compared.

Below, we present the number of timesteps per seconds for the various simulations and for various devices. All simulations were run in the micro-canonical ensemble (NVE), but the number of particles in the simulations vary.

CPU	Computer	Cores	Freq.	LJ	SW	TIP4P	SiC	SiO2
Intel Core i7-4500U	maxwell	4	1.80GHz	357.790	14.694	2.701		0.055
Intel Xeon E5-2620	bigfacet	16	2.10GHz	1415.721	69.960		1.177	0.250
Intel Xeon E5-2670	egil	16	2.60GHz	1580.308	65.556	13.210	1.792	0.269
AMD EPYC 7252	hugefacet	16	3.10GHz	2011.210	117.416	19.943	1.995	0.441
AMD EPYC 7762	griffith	128	2.45GHz	3837.572				5.194
Intel Xeon W-2295	rahman	18	3.00GHz	3134.155	225.875	34.796	3.171	0.712
Apple Silicon M1	laptop	8	3.20GHz	1001.461	75.730			0.399

GPU	Computer	Cores	Freq.	LJ	SW	TIP4P	SiC	SiO2
Nvidia RTX 2070 Super	rahman	2,560	1.60GHz	3103.141	207.475	-	4.169	memerr
Nvidia RTX 2080 Super	bioai	3,072	1.65GHz			-	4.598	memerr
Nvidia RTX 3090	analyze	10,496	1.40GHz	4360.802	363.375	-	8.576	2.165
Nvidia P100-PCIE-16GB	bigfacet	3,584	1.20GHz	4257.105	432.307	-	11.359	2.487
Nvidia A100-PCIE-40GB	hugefacet	8,192	1.40GHz	4820.603	962.502	-	53.733	11.637

Simulation details

All simulations above were run with the same LAMMPS version (February 17th, 2022) to make them reasonably comparable. LAMMPS was built with the packages granular, KOKKOS, KSPACE, manybody, molecular, python and rigid. For the CPU simulations, the command

mpirun -n N lmp -in in.lammps

was run on all cores N on a node. The simulations on GPU nodes were run with the KOKKOS package:

mpirun -n 1 lmp -pk kokkos newton on neigh full -k on g 1 -sf kk -in in.lammps

Each system was simulated 10 times, and the presented number is an average of the number of timesteps per seconds for these simulations. To see the spread in simulation time, see each respective simulation directory.

The different systems

The different bulk systems were chosen to spawn out the force-field space. The Lennard-Jones potential is a two-body potential. The TIP4P is a partly bonded force-field. Stillinger-Weber and Vashishta are non-bonded three-body potentials. We investigate different system sizes to challenge the different devices.

Lennard-Jones (LJ)

We simulate Lennard-Jonesium (e.g. Argon) using the Lennard-Jones potential, as first suggested by [A. Rahman][1]. Our initial system is a face-centered cube with 2,916 particles, and is simulated at 1197K (10000 steps). See src/lennardjones/in.lammps for LAMMPS input script.

Stillinger-Weber (SW)

Silicon is simulated using the Stillinger-Weber potential and the original parameterization by [Stillinger and Weber][2]. The system consists of 64,000 particles initialized in crystal structure and is simulated at 1000K (10000 steps). See src/stillingerweber/in.lammps for LAMMPS input script.

TIP4P/2005 (TIP4P)

Water is simulated using the popular TIP4P potential and the universal parameterization by [Abascal and Vega][3], known as TIP4P/2005. 6,000 water molecules (24,000 particles) are initialized on a grid, and simulated at 300K (1000 steps). See src/tip4p/in.lammps for LAMMPS input script.

Vashishta (SiC)

We simulate silicon carbide using the Vashishta potential and the parameterization by [Vashishta et. al.][4]. The system is initialized with 2,000,000 particles and simulated at 2000K (1000 steps). See src/sic/in.lammps for LAMMPS input script.

Vashishta (SiO2)

Finally, we simulate quartz using the Vashishta potential and the initial parameterization by [Vashishta et al.][5]. The system is initialized with 3,000,000 SiO2 molecules (9,000,000 particles), and simulated at 300K (100 steps). See src/sio2/in.lammps for LAMMPS input script.

Devices

The devices were not, by any rule, picked with care and they do not spawn out the device space in any possible way. In fact, the devices presented here are just the ones that were available when doing the comparison.

References

[1]: A. Rahman, Phys. Rev. A, 136, 405 (1964)
[2]: F. H. Stillinger and T. A. Weber, Phys. Rev. B, 31, 5262 (1985)
[3]: J. L. F. Abascala) and C. Vega, J. Chem. Phys., 123, 234505 (2005) [4]: P. Vashishta, R. K. Kalia and A. Nakano, J. Appl. Phys., 101, 103515 (2007) [5]: P. Vashishta, R. K. Kalia and J. P. Rino, Phys. Rev. B, 41, 12197 (1990)

evenmn / lammps-bulk-benchmark