GROMACS 2022.4 release notes#
This version was released on November 16th, 2022. These release notes document the changes that have taken place in GROMACS since the previous 2022.3 version, to fix known issues. It also incorporates all fixes made in version 2021.6 and earlier, which you can find described in the Release notes.
Fixes where mdrun could behave incorrectly#
Incorrect foreign energy differences for 1-4 interactions on a GPU#
When running free-energy calculations using a GPU without domain decomposition,
the foreign energy differences for 1-4 interactions would be incorrect
when the atoms involved had only charges perturbed and not the atom types.
This issue did not affect free-energy calculations using
This issue did not affect dV/dlambda. If you have used BAR or AWH for free-energy
calculations with atoms with charge manually perturbed in the topoogy and
non-perturbed atom type, we suggest to rerun those with the corrected code
to check if your results were affected by this issue.
The deform option caused parallel mdrun to exit#
With the deform mdp option mdrun would exit with an MPI error when invoked with more than one MPI rank.
Pulling with average force output made mdrun exit at checkpoint writing#
pull_fout_average was set to
yes, gmx mdrun would exit with
an assertion failure when trying to write a checkpoint file.
AMD RDNA devices are now properly marked as “unsupported” with OpenCL#
AMD RDNA GPUs (Radeon RX 5000, 6000, and 7000 series) never worked correctly with OpenCL, usually with a simulation crashing quickly. We now properly mark those devices as incompatible.
AMD GCN (e.g., RX Vega 64) and CDNA/CDNA2 (e.g., Instinct MI100) devices are supported.
Programs reading a tpr file of a large system would exit with random errors#
The tpr file writing and reading code contained a bug that caused reading of systems with more than 100 million atoms to exit with random error messages.
With flexible constraint grompp and mdrun exited with an assertion failure#
Corrected gmx awh friction metric legend#
The output of friction metric with gmx awh -more is the sqrt of
the metric while
sqrt was missing from the legend. This has been
added now. Note that the output of gmx awh -fric, correctly,
does not involve sqrt.
Fixes that affect portability#
Fixed argument checking with nvcc#
Incorrect configure-time checks of flags passed to nvcc resulted in several performance-related flags never being used. Simulations using Nvidia GPUs were correct but may not have achieved best possible performance.
Added compilation support for new CUDA architectures#
The list of NVIDIA CUDA architectures for which code is directly generated for, with a default cmake configuration, and when support exists in the compiler in use, has been updated to include the latest Ada Lovelace and Hopper architectures.