GROMACS

GROMACS features one of the most aggressively optimized GPU-accelerated MD backends in open-source software. It offloads the computationally dominant non-bonded force calculations—specifically the short-range Lennard-Jones potential and real-space Coulombic interactions—directly to CUDA or SYCL kernels.

• Task Parallelism: Dynamically balances work between CPU and GPU cores.
• Dual-GPU Support: Scales efficiently across multiple accelerators within a single node.
• Performance: Routinely achieves microseconds-per-day simulation throughput for systems containing millions of atoms.

To accurately model long-range interactions in periodic systems, GROMACS implements the Particle Mesh Ewald (PME) algorithm. This method splits the Coulombic sum into a rapidly decaying real-space term and a smooth reciprocal-space term solved via Fast Fourier Transforms (FFTs).

• Twin-Range Cutoffs: Uses distinct neighbor lists for short-range van der Waals and electrostatic forces.
• Optimized FFTs: Leverages highly tuned libraries like FFTW and cuFFT for reciprocal-space calculations.
• Verlet Scheme: Employs a buffered Verlet neighbor list to maintain energy conservation while minimizing kernel launches.

GROMACS removes the fastest vibrational degrees of freedom—typically hydrogen-heavy atom bonds—using constraint solvers. This allows the integration time step to be increased from 0.5 fs to 2.0 fs or even 4.0 fs with virtual sites.

• LINCS Algorithm: A linear constraint solver that resets bond lengths iteratively, preferred for its speed and parallelizability.
• SHAKE Algorithm: An older, robust method available for specific legacy force field requirements.
• Virtual Interaction Sites: Replaces hydrogen mass with massless interaction points to further smooth the potential surface.

GROMACS achieves scalability from laptops to supercomputers through a hybrid MPI + OpenMP parallelization strategy. Domain decomposition splits the simulation box spatially across MPI ranks, while OpenMP threads handle the force calculations within each domain.

• Dynamic Load Balancing: Automatically adjusts domain boundaries to equalize computational load across nodes.
• Neutral Territory Method: A specific domain decomposition variant that minimizes communication overhead for non-bonded forces.
• PME Node Dedication: Allows dedicating specific MPI ranks solely to solving the reciprocal-space PME calculation, overlapping communication with computation.

GROMACS provides a native, highly optimized implementation of alchemical free energy calculations. It supports the gradual morphing of one molecule into another via a coupling parameter lambda (λ) to compute relative binding or solvation free energies.

• Multistate Bennett Acceptance Ratio (MBAR): Integrated tools for statistically optimal analysis of alchemical data.
• Soft-Core Potentials: Prevents singularities and numerical instability when atoms appear or disappear at the end states.
• Non-Equilibrium Methods: Supports fast growth and Jarzynski Equality-based protocols for rapid free energy estimates.

To escape kinetic traps and observe rare events, GROMACS integrates multiple enhanced sampling techniques directly into its core engine via a flexible plugin architecture.

• Metadynamics: Supports well-tempered and multiple-walker variants to fill free energy minima along user-defined collective variables.
• Umbrella Sampling: Native pull code for harmonic restraints along a reaction coordinate, with integrated WHAM analysis.
• Replica Exchange MD: Supports temperature, Hamiltonian, and solute-tempering replica exchange schemes across parallel simulations.

The default electrostatic solver in GROMACS for periodic systems. PME splits the Coulombic sum into a short-range real-space term and a long-range reciprocal-space term solved via Fast Fourier Transforms (FFTs). This reduces the computational complexity of long-range interactions from O(N²) to O(N log N), making it feasible to simulate large solvated biomolecular systems with high accuracy.

GROMACS is a leading engine for computing relative binding free energies between ligands. By simulating a non-physical pathway of intermediate states where one molecule is gradually morphed into another, the free energy difference is rigorously calculated using the Multistate Bennett Acceptance Ratio (MBAR) estimator. This is a critical tool for lead optimization in drug discovery.

GROMACS achieves extreme performance by offloading the most compute-intensive calculations to GPUs. The non-bonded force calculations, which scale quadratically with system size, are executed on NVIDIA (CUDA) and Intel (SYCL) accelerators. This heterogeneous parallelization allows a single workstation to achieve simulation throughput that once required a dedicated supercomputing cluster.