- Update to 2023
* Highlights:
- 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).
- 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.
* See https://manual.gromacs.org/2023/release-notes/index.html
- Update to 2022.6
* Bugfix release, see
https://manual.gromacs.org/2023/release-notes/2022/2022.6.html
- Update to 2022.5
* Bugfix release, see
https://manual.gromacs.org/2023/release-notes/2022/2022.5.html