Performance improvements

Up to a factor 2.5 speed-up of the non-bonded free-energy kernel

The non-bonded free-energy kernel is a factor 2.5 faster with non-zero A and B states and a factor 1.5 with one zero state. This especially improves the run performance when non-perturbed non-bondeds are offloaded to a GPU. In that case the PME-mesh calculation now always takes the most CPU time.

Proper dihedrals of Fourier type and improper dihedrals of preriodic type are SIMD accelerated

Avoid configuring the own-FFTW with AVX512 enabled when GROMACS does not use AVX512

Previously if GROMACS was configured to use any AVX flavor, the internally built FFTW would be configured to also contain AVX512 kernels. This could cause performance loss if the (often noisy) FFTW auto-tuner picks an AVX512 kernel in a run that otherwise only uses AVX/AVX2 which could run at higher CPU clocks without AVX512 clock speed limitation. Now AVX512 is only used for the internal FFTW if GROMACS is also configured with the same SIMD flavor.

Update and constraints can run on a GPU

For standard simulations (see the user guide for more details), update and constraints can be offloaded to a GPU with CUDA. Thus all compute intensive parts of a simulation can be offloaded, which provides better performance when using a fast GPU combined with a slow CPU. By default, update will run on the CPU, to use GPU in single rank simulations, one can use new ‘-update gpu’ command line option. For use with domain decomposition, please see below.

GPU Direct Communications

When running on multiple GPUs with CUDA, communication operations can now be performed directly between GPU memory spaces (automatically routed, including via NVLink where available). This behaviour is not yet enabled by default: the new codepaths have been verified by the standard GROMACS regression tests, but (at the time of release) still lack substantial “real-world” testing. They can be enabled by setting the following environment variables to any non-NULL value in your shell: GMX_GPU_DD_COMMS (for halo exchange communications between PP tasks); GMX_GPU_PME_PP_COMMS (for communications between PME and PP tasks); GMX_FORCE_UPDATE_DEFAULT_GPU can also be set in order to combine with the new GPU update feature (above). The combination of these will (for many common simulations) keep data resident on the GPU across most timesteps, avoiding expensive data transfers. Note that these currently require GROMACS to be built with its internal thread-MPI library rather than any external MPI library, and are limited to a single compute node. We stress that users should carefully verify results against the default path, and any reported issues will be gratefully received to help us mature the software.

Bonded kernels on GPU have been fused

Instead of launching one GPU kernel for each listed interaction type there is now one GPU kernel that handles all listed interactions. This improves the performance when running bonded calculations on a GPU.

Delay for ramp-up added to PP-PME tuning

Modern CPUs and GPUs can take a few seconds to ramp up their clock speeds. Therefore the PP-PME load balancing now starts after 5 seconds instead of after a few MD steps. This avoids sub-optimal performance settings.