Highlights#

GROMACS 2023 was released on February 6th, 2023. Patch releases may have been made since then, please use the updated versions! Here are some highlights of what you can expect, along with more detail in the links below!

As always, we’ve got several useful performance improvements, with or without GPUs, all enabled and automated by default. In addition, several new features are available for running simulations. We are extremely interested in your feedback on how well the new release works on your simulations and hardware. The new features are:

  • The SYCL GPU implementation, which is the GPU portability layer that supports all major GPU platforms, has received major extensions in support for both platforms and features. To ensure portability in practice, the GROMACS GPU portability layer is actively developed with multiple SYCL implementations (hipSYCL, oneAPI DPC++, IntelLLVM) and regularly tested on multiple GPU backends.

    • SYCL supports more GPU offload features: bonded forces and direct GPU-GPU communication with GPU-aware MPI.

    • SYCL hardware support includes AMD (including RDNA support added here) and Intel for production as well as NVIDIA GPUs (not for production).

    • SYCL optimizations targeting important HPC platforms.

  • PME decomposition has been optimized and extended to support offloading the entire PME calculation to multiple GPUs, including the FFT computation; when combined with cuFFTmp or heFFTe this enables much improved strong scaling (experimental feature).

  • CUDA Graph support has been added to execute GPU-resident single-/multi-GPU simulations using thread-MPI entirely on the GPU to improve performance (experimental feature).

  • Apple M1/M2 GPUs are now supported via the OpenCL GPU backend.

  • New ensemble temperature mdp options allow setting the temperature of the ensemble for simulations without temperature coupling or with different reference temperatures.

  • With gmx dssp, GROMACS now has a native implementation of the DSSP algorithm, which replaces gmx do_dssp.