GROMACS version:
GROMACS modification: Yes/No
Hi i have completed 2 simulation but both of them didn’t use 100% gpu. I have a laptop with rtx 4060 mobile and i5 12500h processor my first simulation with rtx 4060 consuming 80% to 85% and 8 threads of cpu around 60 to 65% consumption but when i read the log file it showed 45% gpu wait time. So i ran another simulation ( this one of 200ns) but with 14 thread of cpu consuming 99 to 100% of cpu and gpu around 50 to 55% and this time the gpu wait time was reduced to 16.5%. So is there any way to use 100% of my gpu and also get reduced gpu wait time.
Hi!
What is the size of your system? E.g., if you have 20 000 atoms, it’s unlikely you will be able to use 100% of the GPU, the system is just too small.
Be advised that GPU utilization and GPU Wait time (%) are not related the way you expect. Higher % of “GPU wait time” means that CPU is waiting for the GPU, i.e., the GPU has more work than CPU (which is usually desirable), i.e. that GPU is more fully utilized.
Furthermore: GPU utilization reported by the driver is a rather meaningless metric by itself: in some cases, 100% is not achievable (small systems), and in others, high utilization can lead to worse performance (ns/day) than low utilization.
I suggest reading the “Running mdrun with GPUs” and “Checking and improving performance” in our performance guide to get a general overview of how to tune GROMACS performance.