Date: Tue, 04 Feb 2025 14:39:50 +0800 (CST)
Dear users and developers,
I am trying to use random acceleration MD (RAMD) simulation on GPU. A ligand-bound HSP90 system (PDB: 2VCI) was used for all tests. After equilibration, RAMD simulation was run in NTP conditions (ntb=2, npt=1, ntt=3, gamma_ln=2.0, temp0=300.0) with the following RAMD parameters: ramdboost=0, ramdboostfreq=1, ramdboostrate=0, ramdint=1. This parameter combination is expected to trigger ramd module while applying zero perturbation to the system, as the additional acceleration remains zero.
When I run RAMD on GPU, as mentioned above, I noticed an unexpected temperature fluctuation.
Can anyone confirm whether anyone also faced the same issue with temperature fluctuations?
After investigating further by reading the code, I suspect that it may be related to the calling of the subroutine gpu_download_frc(frc) (line 232, src/pmemd/src/ramd.F90).
If my limited understanding is correct, when this subroutine is called in an NPT system, the returned forces include both bonded and non-bonded (NB) contributions (as seen in lines 2398-2402 of src/pmemd/src/cuda/gpu.cpp). Consequently, when a random perturbation force is added to frc and uploaded to the GPU (line 246, src/pmemd/src/ramd.F90), and ultimately used to update velocity (lines 86-92, src/pmemd/src/cuda/kU.h), the NB force component may be double-counted, leading to incorrect velocity updates and the observed temperature increase.
Based on this understanding, I think one could avoid updating the force which is obtained using gpu_download_frc(frc). Instead, only updating the extra force with inherent subroutine gpu_upload_frc_add(frc) (lines 2304-2333, src/pmemd/src/cuda/gpu.cpp) could bypass this issue. Hope I am thinking in a right direction.
For your convenience, I have prepared a patch file to address this issue. On applying the patch the issue can be fixed. I am not sure if this is correct fix. I would appreciate it if someone from developers could review the patch and provide feedback/insights.
Thank you,
Han CHENG
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
(image/png attachment: temp.png)
- text/plain attachment: ramd.patch
