« first day (325 days earlier)      last day (837 days later) » 

4:46 PM
10
A: How to calculate the computational time required for a MD simulation

PBHThe simulation speed and efficiency depends on the hardware and software that you use. Things that you should keep in mind are, Every computer is unique in its ability. So only someone with access to the exact same hardware (processor, gpu etc.) can test this out and tell you. As far as I know, ...

Small addition to the (good) advice about making your own scaling graphs: Since scaling is typically limited by system size, sometimes you can double performance by going from 1 GPU to 2 GPUs, but then get zero improvement going from 2 GPUs to 4 GPUs. It's always good to explicitly check performance when running at scale. — dwhswenson yesterday
Exactly. You always have to try and see to determine the performance scaling. Its either trial and error or experience. Also keep in mind the precision of the calculations. Sometimes single and mixed precision may yield similar results to double precision but with much higher calculation speeds. — PBH yesterday
@PBH Thanks. In your personal experience (your own projects), roughly speaking, how long does it take? — scamander yesterday
@scamander as I said before, my projects cannot be compared with yours. But for your reference, I work with solid Si/Ge nano structures of around 100k to 300k atoms in each calculation. In the two workstations I own, the computational speeds vary from 2-8 hours/ns where one is accelerated using CUDA while the other uses pure CPU for calculations. — PBH yesterday
For my calculations, I vary my core count from 2 to 32 CPU cores and the GPU acceleration is achieved by an NVIDIA RTX3080 and an RTX2070 Super. In these GPUs, the fp16 (single precision) performance is significantly higher than fp64 double precision and so doing a calculation in single precision is much faster, and the difference in results in negligible. So check the performance in single, mixed as well as double precision modes (which is software dependent and I cannot comment on how it is done within Gromacs) while checking the accuracy of the results as you continue. — PBH yesterday
 

« first day (325 days earlier)      last day (837 days later) »