GROMACS 2024.3 release notes#
This version was released on August 29th, 2024. These release notes document the changes that have taken place in GROMACS since the previous 2024.2 version, to fix known issues. It also incorporates all fixes made in version 2023.5 and earlier, which you can find described in the Release notes.
Fixes where mdrun could behave incorrectly#
Remove double dispersion correction with MTTK barostat#
When the MTTK barostat was used with dispersion correction, the pressure correction was applied twice. This caused the density to be too high.
Correct zero Verlet buffer for system with only large masses#
For system with most masses much larger than zero, the Verlet buffer estimate could be zero. Coarse-grained simulations could be affected by this issue.
Fix incorrect energy group pair assignment of pair interactions#
Energies for bonded pair interactions, typically 1-4 interactions, would be assigned to incorrect energy group pairs. With free-energy calculations this could lead to illegal memory access.
Correct AWH metric with free-energy and Beutler soft-core#
The AWH metric was incorrect for free-energy dimensions when Beutler soft-core interactions were used. The free-energies themselves were unaffected by this issue.
Implement Lennard-Jones force-switch in free-energy kernel#
Perturbed non-bonded Lennard-Jones interactions would use plain cut-off with potential shift when force-switch was specified. The errors this causes on the potential and force are negligible with the typical Charmm setup. When (de)coupling a complete, dense system there is a measurable error in the pressure, but this is again negligible when coupling a single molecule.
mdrun now computes the initial bonded distances with periodic molecules#
When using periodic molecules, mdrun could quit with an error message about multiple assignments of bonded interactions telling the user to report the issue. The actual source of this was that mdrun did not determine the communication distance required for bonded interactions. Now this distance is determined.
Avoid using invalid device streams#
Issue has only been seen via assertions in the GCC 13 standard library that guard against null-pointer dereferences. The dereferenced null pointers were never used in actual simulations.
Fix masks and alignment checks in Arm SVE SIMD#
The double precision SVE implementation contained an incorrect mask for SVE implementations of 256 bits or more. This would lead to instantaneous extreme forces/crashes on the first step, so it should never have influenced any production simulations unnoticed. The alignment checks for some SVE routines have also been modified to correspond to the SIMD width, although this is formally not required on Arm, so it will not have caused any errors.
Added an assert to ensure valid cutoff when calculating atom density#
Simulations with rvdw and rcoulomb set to 0 could crash
with a segmentation fault when calculating the effective atom density.
This is now handled by asserting valid values and halting with a proper
explanation.
Fixes for gmx tools#
Permit gmx tune_pme to understand md.log files#
The format of md.log files changed in 2024 which broke the ability of
gmx tune_pme to understand how previous runs worked by parsing that
file. Now gmx tune_pme understands better.
Avoid crash of convert-tpr on velocity-less tpr files#
The convert-tpr now works on tpr files that do not contain velocities, as generated for, e.g., energy minimization.
Avoid premature termination when using gmx hbond#
Fixed an error that caused gmx hbond to prematurely terminate if there were no donors or acceptors of hydrogen bonds in the chosen selection.
Add grompp warning about zero AWH cover diameter when sharing a bias#
Fix TPR generation with Andersen Massive thermostat and constraints#
A restriction to not generate TPR files for the Andersen thermostat with constraints was incorrectly applied to also the Andersen Massive thermostat.
Fixed filenames of generated CP2K files in QMMM MdModule#
If there were no files provided with grompp -qmi option, module used topol/_cp2k.inp instead of topol_cp2k.inp
Added a check in gmx grompp to ensure that cutoffs are valid#
At least one of rvdw and rcoulomb must be > 0 when using the Verlet cutoff scheme.
This is now checked by gmx grompp.
Fix gmx dump with options -sys and -orgir#
These gmx dump options had no effect. This has now been fixed.
Fixes that affect portability#
Fix build when GoogleTest is installed system-wide#
In some cases (e.g., on FreeBSD), GROMACS build system can mix up different versions of GoogleTest leading to compilation errors. Now we always prefer the bundled version.
Miscellaneous#
Fix detection of rocFFT installed in non-standard locations#
Fix performance regression on PVC with oneAPI 2024.2#
oneAPI 2024.2 enabled automatic GRF selection by the device compiler, causing up to 50% worse NBNxM kernel performance on Intel Data Center GPU Max (Ponte Vecchio) devices. We now force the use of small GRF mode on PVC to avoid that. Arc GPUs were not affected.