GROMACS modification: No
Here post your question
I am interested in knowing how much of the performance of my runs come from the CPU cores, and how much from the GPU. In my case, the PME is always run on the CPU and I can see in md.log things like
Would this mean that the CPU time is roughly 63% of the total? If so, how much % of the total performance [ns/day] would be generated by CPU?
These breakdowns are unfortunately a bit hard to interpret, it’s not as simple as CPU taking x amount of the simulation and GPU taking y amount, because CPU and GPU aspects happen concurrently and it’s hard to balance them so that they finish at the same time.
Here, you can see the “Wait GPU NB local/nonlocal” contribute very little - that means that the GPU step (nonbonded forces) typically finishes before the CPU tasks on the same step. So, the PME calculations are taking up 63% of your step time, and the GPU calculations are taking some unknown amount of time to complete, where all we can say is that it finishes sometime during that large chunk of PME computation.