Collagen GROMACS Benchmark

GPU benchmark and simulation inputs for a collagen microfibril at coarse-grained (CG) and all-atom (AA) resolution. The benchmarks were run with GROMACS 2026.0 on an NVIDIA L40S and an AMD EPYC 7413 host.

What was benchmarked

The project contains two related studies:

Directory	System	Purpose
cg_constant_pull/	Martini 3 + Go CG, 242,794 particles, 24 groups pulled at 1000 pN	Compare CPU/CUDA and thread/offload configurations
cg_aa_no_pull/	Unforced CG and AA versions of the same microfibril	Compare CG with AA and water-in with water-out trajectory output

The .tpr files are compiled GROMACS run inputs. The .mdp files in cg_aa_no_pull/ record the no-pull run parameters. Archived .log files are retained as the source for the reported performance numbers; large trajectories and disposable benchmark outputs were removed.

Main results

Benchmark	Configuration	Throughput
CG 1000 pN pull, CPU fallback	1 MPI × 20 OMP	29.55 ns/day
CG 1000 pN pull, CUDA	1 MPI × 20 OMP, nonbonded on GPU	118.93 ns/day
CG 1000 pN pull, fastest CUDA	1 MPI × 8 OMP, nonbonded + bonded on GPU	270.81 ns/day
CG no pull, water in trajectory	fastest CUDA config	285.77 ns/day
CG no pull, water out of trajectory	fastest CUDA config	287.24 ns/day
AA no pull, water in trajectory	1 MPI × 8 OMP, PP + PME on GPU	15.36 ns/day
AA no pull, water out of trajectory	1 MPI × 8 OMP, PP + PME on GPU	16.06 ns/day

The fastest pull configuration is about 9.2× faster than the 20-thread CPU fallback. For the no-pull system, CG produces about 18.6× more simulated time per day than AA. Removing pulling improves CG throughput by about 5.5%.

The 20 fs CG result (957.93 ns/day) is shown separately as experimental: an earlier run encountered LINCS constraint failures, so it is not a validated production configuration.

Recommended configurations

For CG on this host:

-ntmpi 1 -ntomp 8 -gpu_id 0 -nb gpu -bonded gpu -rdd 3.9

For AA:

-ntmpi 1 -ntomp 8 -rdd 3.9

Use the gromacs-cuda Conda environment. The installed OpenCL build did not detect the L40S and therefore ran on CPU.

See the directory documentation for exact commands:

• cg_constant_pull/README.md
• cg_aa_no_pull/README.md

Running the no-pull benchmarks

cd /root/collagen_benchmark/cg_aa_no_pull

./run_sim.sh cg       # short CG benchmark, water in xtc
./run_sim.sh cg_now   # short CG benchmark, water out of xtc
./run_sim.sh aa       # short AA benchmark, water in xtc
./run_sim.sh aa_now   # short AA benchmark, water out of xtc

For a full production run:

MODE=production ./run_sim.sh cg

noW changes only which atoms are written to the compressed trajectory; water remains present in the simulated system.

Repository storage note

GitHub rejects individual files larger than 100 MB, and Git LFS quota was not available when this repository was created. These all-atom inputs therefore remain local and are ignored by Git:

• cg_aa_no_pull/aa.tpr
• cg_aa_no_pull/aa_noW.tpr
• cg_constant_pull/md.tpr

The smaller CG .tpr inputs, MDP files, documentation, scripts, and benchmark logs are included in the repository.