[gmx-developers] Wrong Free Energy on GPU

[Igor Leontyev]

Dear gmx developers,


Simulation of the Relative Free Energy on GPUs (using gmx 2020-2022.3) gives dramatically different results between BAR and TI (Thermodynamic Integration) methods. The problem is observed in all tested mutations even in the simplest 1-atom mutation of Acetone->Acetamide in Water with PME:


GPU results: BAR -173.70 +/- 0.12 TI -181.7074 +/- 0.1501

cpu results: BAR -181.42 +/- 0.10 TI -181.5248 +/- 0.1512


As you can see the BAR results disagree between GPU and cpu computations while TI results are in the perfect agreement.

Sincerely,
Igor Leontyev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20221014/8cd1161a/attachment.html>