Performance improvements

Free-energy kernels on GPU with CUDA for perturbed non-bonded interactions

Free-energy calculations can now run faster on hardware with CUDA GPUs. The benefits are largest when the CPU is relatively weak and/or there are many perturbed non-bonded interactions. This feature has been contributed by “MetaX Integrated Circuits Technology”. Note that this feature is currently marked as experimental.

Optimized PME force transfers for multi-GPU runs with GPU-aware MPI

When using a GPU build with an (external) GPU-aware MPI library, the transfer of PME force data between GPUs has been optimized to use non-blocking MPI calls, improving performance by up to around 5%.

More permissive AWH histogram equilibration check

The AWH histogram equilibration check, which is recommended to be used with multiple walkers, used a too strict tolerance of 20% maximum deviation. This tolerance can now be set by the user and the default value is 30%. This value provide significantly faster convergence when using multiple walkers. Since it does not deteriorate the convergence with a single walkers, the histogram equilibration check is now turned on by default.