===== Benchmarks ==

** Runtime/Scaling behaviour on GPU platforms **

The wall clock time of jobs is influenced by following factors:

  - **Scaling with respect to each GPU**: the number of threads (SMP) and/or processes (MPI,hybrid) determines the load on each GPU card
  - **Scaling with the number of GPU**: Only if the job is big enough and exploits the full power of at least one GPU card
  - Binding of threads/processes to specific CPU cores slows down and wastes CPU cycles


** Memory considerations **

  * SMP runs of NAMD only use about 600 MB of resident memory