GROMACS 2023.5 release notes#
This version was released on May 3rd, 2024. These release notes document the changes that have taken place in GROMACS since the previous 2023.4 version, to fix known issues. It also incorporates all fixes made in version 2022.6 and earlier, which you can find described in the Release notes.
Fixes where mdrun could behave incorrectly#
The Verlet buffer estimate could be wrong with initial coordinates out of the box#
Due to incorrect periodic boundary treatment in the density estimate for the Verlet buffer calculation, the Verlet buffer could be off when many atoms were outside the rectangular unit cell. In nearly all cases this would lead to a slightly too large buffer and almost never to a too short buffer.
With energy minimization, dispersion correction energies were double counted#
The only affected the reported energies, as dispersion correction does not affect the forces on the atoms.
Correct forces with virtual_sites1#
Virtual sites with a single constructing atom did not have their force distributed to the constructing atom. This likely did not cause silent errors, as this completely defeats the purpose of such virtual sites.
Avoid rounding errors with Parrinello-Rahman coupling of triclinic boxes#
With triclinic boxes and isotropic or diagonal Parrinello-Rahman pressure coupling, the off-diagonal scaling matrix arguments could be non-zero. This was harmless, apart from causing an assertion failure when update was executed on a GPU.
Add checks for the combination of multiple time stepping and SD integrator#
The documentation says that this combination is not supported, but the check for this in grompp was lost. Now both grompp and mdrun will exit with an error.
Correct dH/dlambda with MTS and perturbed constraints#
When doing free-energy calculations with multiple time stepping and perturbed constraint lengths, the slow force contributions to dH/dlambda were incorrectly multiplied by the MTS factor. The effect of this is usually rather small when only integrating the PME grid part less frequently, but can be substantial when the non-bonded pair interactions are not integrated at every step.
Fixes for gmx tools#
Fix grompp crash with atoms far out of the box#
The Verlet buffer tolerance code did not put atoms correctly in the box, which could lead to illegal memory access in grompp
Fix LJ combination rules mixup in nonbonded-benchmark tool#
The gmx nonbonded-benchmark tool would use geometric LH combination rules instead of Lorentz-Berthelot and Lorentz-Berthelot instead of no combination rules.
Fix make_ndx behavior with splitres#
Since GROMACS 2023, gmx make_ndx would only output the first atom
in each residue when using splitres command. Now the old behavior
is restored, where all atoms in the residue are printed.
Fix invalid memory access in gmx make_ndx when using splitres/splitat#
Could manifest in application crashes or garbled output data (for example, the output group names had empty prefix).
Fixed references to old tool names in some commands#
There were references to old tool names, e.g., g_energy and g_bar in some commands.
Fixes that affect portability#
Use RDTIME instead of RDCYCLE on RISC-V#
Starting with Linux 6.6, the rdcycle instruction cannot be called from
userland for security reasons, making GROMACS fail with SIGILL.
Now it is replaced with rdtime.
Miscellaneous#
Fix crashes with GPU direct communication for some atypical run configurations#
When GPU direct communication was used in combination with atypical run conditions (e.g. many thread-MPI tasks on each GPU), crashes could occur when the tasks became very out of sync. This has now been made more robust.