GROMACS¶

GROMACS is a versatile package to perform molecular dynamics, i.e. simulate the Newtonian equations of motion for systems with hundreds to millions of particles and is a community-driven project.

Package availability¶

Compute node type	Available
CPU	✓
NVIDIA A100 GPU	✓
NVIDIA Grace Hopper GPU	✓

Run GROMACS¶

GROMACS is available in both MPI and non-MPI versions. In single-GPU-node tests, both versions produced comparable results. For multi-node runs, use the MPI version.

To load the non-MPI version:

Bash

module load gcc/13.4.0 gromacs/2026.4

To load the MPI version:

Bash

module load gcc/13.4.0 mpich/5.0.1 gromacs/2026.4

For the examples and benchmarks in this documentation, we use the non-MPI version.

NVIDIA A100 GPU¶

Bash

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --output=gromacs_stdout.txt
#SBATCH --error=gromacs_stderr.txt
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=8
#SBATCH --cluster=gmerlin7
#SBATCH --partition=a100-hourly
#SBATCH --gpus=4
#SBATCH --mem=0
#SBATCH --exclusive

module purge
module use Alps_A100
module load gcc/13.4.0 gromacs/2026.4

gmx mdrun \
    -s /path/to/input/topol.tpr \
    -ntmpi $SLURM_NTASKS_PER_NODE \
    -ntomp $SLURM_CPUS_PER_TASK \
    -nb gpu \
    -pme gpu \
    -npme 1 \
    -update gpu \
    -bonded gpu \
    -pin on \
    -v \
    -gpu_id $CUDA_VISIBLE_DEVICES

Grace Hopper GPU¶

Bash

#!/bin/bash
#SBATCH --time=01:00:00
#SBATCH --output=gromacs_stdout.txt
#SBATCH --error=gromacs_stderr.txt
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=36
#SBATCH --cluster=gmerlin7
#SBATCH --partition=gh-hourly
#SBATCH --gpus=4
#SBATCH --mem=0
#SBATCH --exclusive

module purge
module load gcc/13.4.0 gromacs/2026.4

export GMX_DISABLE_DIRECT_GPU_COMM=1

gmx mdrun \
    -s /path/to/input/topol.tpr \
    -ntmpi $SLURM_NTASKS_PER_NODE \
    -ntomp $SLURM_CPUS_PER_TASK \
    -nb gpu \
    -pme gpu \
    -npme 1 \
    -update gpu \
    -bonded gpu \
    -pin on \
    -v \
    -gpu_id $CUDA_VISIBLE_DEVICES

Resource allocation

Request resources according to the requirements of your job. Requesting more resources than necessary may increase queue wait times and reduce cluster efficiency.

Benchmark results¶

We performed a single-node performance benchmark using the Satellite Tobacco Mosaic Virus (STMV) benchmark, comparing performance on A100 and Grace Hopper nodes.

For most of the tested configurations, 8 tasks per node provided the best performance and appears to be the optimal configuration for single-node runs. For a single GPU, 4 tasks per node has a slight advantage in performance.

Run	Version	GPUs	CPU (Threads)	N tasks	Memory (Gb)	Performance (ns/day)
STMV	2026.3	1 A100	8 per task	2	118	21.9
STMV	2026.3	1 A100	4 per task	4	118	23.0
STMV	2026.3	1 A100	2 per task	8	118	22.8
STMV	2026.3	1 GH	36 per task	2	118	44.1
STMV	2026.3	1 GH	18 per task	4	118	45.9
STMV	2026.3	1 GH	9 per task	8	118	44.7

STMV	2026.3	2 A100	16 per task	2	236	26.0
STMV	2026.3	2 A100	8 per task	4	236	39.3
STMV	2026.3	2 A100	4 per task	8	236	44.8
STMV	2026.3	2 A100	2 per task	16	236	41.3
STMV	2026.3	2 GH	36 per task	4	236	75.4
STMV	2026.3	2 GH	18 per task	8	236	85.2

STMV	2026.3	4 A100	16 per task	4	472	73.8
STMV	2026.3	4 A100	8 per task	8	472	85.7
STMV	2026.3	4 A100	4 per task	16	472	74.6
STMV	2026.3	4 GH	72 per task	4	472	137.1
STMV	2026.3	4 GH	36 per task	8	472	153.4
STMV	2026.3	4 GH	18 per task	16	472	128.2