Known issues relevant for developers¶
This is a non-exhaustive list of known issues that have been observed and can be of interest for developers. These have not been solved because they are either outside the scope of the GROMACS project or are are simply too difficult or tedious to address ourselves.
Issues with GPU timer with OpenCL¶
When building using OpenCL in Debug mode, it can happen that the GPU timer state gets
corrupted, leading to an assertion failure during the mdrun.
This seems to be related to the load of other, unrelated tasks on the GPU.
GPU emulation does not work¶
The non-bonded GPU emulation mode does not work, at least for builds with GPU support; then a GPU setup call is called. Also dynamic pruning needs to be implemented for GPU emulation.
OpenCL on NVIDIA Volta and later broken¶
The OpenCL code produces incorrect results on Volta and Turing GPU architectures from NVIDIA (CC 7.0 and 7.5). This is an issue that affects certain flavors of the nonboded kernels, most likely a result of miscompilation, and there is no known workaround.
PME decomposition automated task assignment broken¶
When there are two or more ranks on a node doing combined PP and PME
work (i.e no separate PME ranks) and more GPUs are detected than
ranks, the automated task assignment fails and GROMACS aborts with
“Error in user input” message. You can work around
this by using -gpu_id or GMX_GPU_ID or limiting the number of
visible GPUs.