I'm interested in discussing the performance of the new version of NAMD. It seems that in this updated version, there's an increased reliance on GPU processing. However, despite using this new version, I haven't observed any notable improvements compared to the previous iteration.

Below are the submission codes and the average processing times for both versions:

NAMD_3.0b6_Linux-x86_64-multicore-CUDA/namd3 +p24 +idlepoll $NAMD_INPUT > $NAMD_OUTPUT
Average time = 93.65 ns/day

NAMD_3.0b3_Linux-x86_64-multicore-CUDA/namd3 +p24 +idlepoll $NAMD_INPUT > $NAMD_OUTPUT
Average time = 93.55 ns/day

I don't know if there is some wrong. What do you think?

  • $\begingroup$ Did the provided answer solve your problem? If yes, mark the answer as accepted. $\endgroup$ Commented Apr 3 at 15:40
  • $\begingroup$ No, It didn't. I will write the answer. $\endgroup$ Commented Apr 4 at 10:48

2 Answers 2


The performance of NAMD is highly dependent on the underlying system hardware and the settings that you had used during the compilation process. However, here are a few things that you can check:

  1. Try using +isomalloc_sync in your script along with +idlepoll.

  2. Ensure that you have the correct MPI bindings, which means that you might have to recompile the binary.

  3. Check if you are having full utilization of the available GPU and CPU cores.

  4. Most important, please check if you are using reasonable values for your time step and write out frequencies.


In order to harness the advantages of GPU-resident simulation mode and achieve a performance boost, follow these steps:

1.Include the following keyword in the input file to activate the mode:

CUDASOAintegrate on
  1. Additionally, ensure to include the following line in the submission script:


By implementing these configurations, you can double the performance while utilizing GPU-resident simulation mode.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .