Dear experts
I would like to initiate a technical discussion regarding GPU acceleration in GROMACS, particularly for energy minimization (EM) steps. My current system involves a large protein complex of approximately 2000 proteins, and I am exploring ways to optimize performance. While I understand that EM does not fully support PME offloading to GPUs, I am interested in discussing potential strategies to accelerate this stage (e.g., short-range GPU usage, tuning nstlist, or CPU–GPU balancing) and how best to transition into GPU-accelerated NVT/NPT/MD runs for such a large system.
I have read that Energy minimization step cannot be accelerated using GPUs (Need to read more on it) unlike NVT, NPT and MD production run.
Looking forward for a fruitful discussion.
Yes, in recent GROMACS versions (2024+), energy minimization (EM) can partially leverage GPUs for force calculations like short-range non-bonded interactions and PME offloading, providing some acceleration, though the core minimization algorithms (steepest descent, conjugate gradient, or l-bfgs) remain CPU-bound and sequential. Full GPU residency isn't supported for EM, limiting gains compared to MD, but you can optimize by enabling GPU for compatible force terms (e.g., via -nb gpu -pme gpu in mdrun), tuning nstlist to 100-300 for reduced neighbor searches (balancing accuracy and speed), and using CPU-GPU balancing with fewer MPI ranks per GPU (e.g., 1-2) to minimize data transfer overhead. For your ~2000-residue system, start with l-bfgs for faster convergence than steepest descent, and monitor performance with gmx tune_pme to auto-optimize PP/PME load.
To transition to GPU-accelerated NVT/NPT/MD: After EM converges (check max force
- Badges
- Science topic
- Similar topics
- Higher Education
- Universities