GROMACS version: 2021.4
GROMACS modification: Yes
Hello everyone,
I was trying to run Gromacs (patched with plumed 2.8) on GPU, but I am noticing that using GPU for calculations is actually reducing performance when compared to running the simulation on CPU only.
The configuration of my computer is:
CPU: AMD Threadripper PRO 3975WX
GPU: RTX A4000
GCC version: 11.2
CUDA version: 11.7.1
Can anyone help me figure out what is happening?
Thanks
You can experiment with
pme-order
and
ewald-rtol
in your mdp file. I have different computation time with pme order 4, 5, 6 and also different for ewald-rtol 10-4, 10-5 and 10-6.
Are you sure that you installed gromacs with GPU correctly?
https://manual.gromacs.org/documentation/2022/install-guide/index.html
On this step
cmake … -DGMX_BUILD_OWN_FFTW=ON -DREGRESSIONTEST_DOWNLOAD=ON
You must add
-DGMX_GPU=CUDA
so you should use
cmake … -DGMX_BUILD_OWN_FFTW=ON -DREGRESSIONTEST_DOWNLOAD=ON -DGMX_GPU=CUDA
Hello @JakubH
I did switch on the cuda flag while compiling.
I also tried changing pme-order and ewald-rtol which did improve the performance but still GPU offloading is not improving the performance in comparison to CPU only calculations.
I also tried the same simulation with a Ryzen 7 5900HX (16 threads) and RTX 3060m gpu laptop (identical gcc and cuda ) and it is giving me more than double the performance in comparison to my RTX A4000 workstation.
Thanks for your help
