GROMACS currently prevents coordinate updates for virtual sites from running on the GPU in the official 2026.3 release. Investigators reported a modified build (Optimized) that handles OPC water virtual sites inside the GPU update and compared it directly with the Reference build. The two builds located the OPC negative charge site with the same tiny placement error (2.3 x 10^-7 nm) and produced relative force differences below 10^-6. These numeric checks were explicit; the study team observed near-identical low-level arithmetic between builds, indicating that moving the coordinate update to the GPU did not introduce appreciable numerical drift in their tests.
The authors ran five matched pairs of 5-microsecond simulations on the Trp-cage miniprotein to probe whether altered update handling would change folding behavior. The Reference build folded in three of five runs while the Optimized build folded in five of five runs, with a reported P value of 0.44. Investigators reported that the folded ensembles matched between builds. In other words, folding outcomes and the ensemble-level structures did not diverge in these matched, multi-microsecond tests, and the data indicated no systematic bias introduced by the GPU-based coordinate update under the conditions they tested.
OPC water is recommended for the ff19SB protein force field because of distinct bulk properties relative to simpler models. The study team measured density, diffusion coefficient, dielectric constant and the second hydration shell—properties that distinguish OPC from TIP3P—and found both builds reproduced these metrics. Data indicated that moving the virtual-site coordinate update to the GPU preserved the water model’s characteristic thermodynamic and dynamic behavior in the tested systems. The source did not report additional water-model variants beyond OPC and TIP3P in these validation runs.
To probe production-size behaviour, investigators simulated a 166,826-atom membrane system containing the glucagon-like peptide-1 receptor. Residue-level hydration agreed strongly between the two builds, with a reported correlation coefficient r = 0.99. This high correlation suggests that local solvent interactions around protein residues were effectively unchanged by the Optimized update strategy in this membrane-protein context. The authors reported these hydration comparisons as direct, quantitative checks of physical behaviour at scale.
Performance testing on two GPUs showed the Optimized build ran between 1.57 and 1.66 times as fast as a tuned Reference. Investigators noted that 87–97% of the saved runtime derived from moving the coordinate update to the GPU. The implication is clear: the coordinate-update stage—previously a CPU-bound bottleneck when virtual sites were present—became largely GPU-resident, yielding the bulk of the speed-up observed in their multi-GPU runs. The authors provided these performance ratios and the reported fraction of time saved as direct measurements from their benchmarking; they did not extend performance claims to other hardware configurations or to single-GPU setups in the source text.
This report is a preprint and has not been certified by peer review; the authors declared no competing interest. The source provided specific numerical and ensemble-level checks for the systems tested but did not report broader cross-force-field validation, diverse hardware benchmarks beyond the two-GPU tests, or long-term stability beyond the simulations described. Investigators referenced a code repository in the footnotes, but the paper did not include additional systematic tests across other water models or force-field combinations. Open questions that remain in the source include how the Optimized update performs across a wider set of systems, whether other virtual-site uses behave identically, and how performance scales on different GPU generations. These items represent the concrete next steps implicit in the authors’ reported limitations and are left as unanswered by the source.