Hello amber community,
I hope everyone is well.
I am experiencing a poorer than expected performance when running explicit solvated system, with performances ~330 ns/day for a system of size ~18k atoms using amber24 on an Ubuntu 24.02 system. Here is my workstation hardware;
CPU: Intel Xeon W-2295, 18 cores . 3 GHz
GPU: NVIDIA GeForce RTX 3080 10GB
Storage: 1 TB M.2 2280 PCIe NVMe Gen4x4 + 2 TB M.2 2280 PCIe NVMe Gen4x4
RAM: 64GB (4 x 16GB) DDR4 2933 MHz
PSU: 1000W
This is the prod.in using pmemd.cuda;
igb = 0, imin = 0, irest = 1, ntx = 7,
ntb = 2, pres0 = 1.0, ntp = 1,
taup = 2.0,
cut = 10.0, ntr = 0,
ntc = 2, ntf = 2,
tempi = 300.0, temp0 = 300.0,
ntt = 3, gamma_ln = 2.0,
nstlim = 250000000, dt = 0.002,
ntpr = 50000, ntwx = 50000, ntwr = 50000,
tol = 0.00001
/
Troubleshooting I’ve already performed;
* I have checked that the GPU is being utilized, although the memory usage is only 472 Mb. Both the NVIDIA drivers (580.95.05) and CUDA versions (13.0) are up-to-date, and the correct version of CUDA is being used.
* The GPU is in the correct PCIe slot, x16 gen 3, and is forced x16 speed in the BIOS.
* The GPU isn’t thermal throttling, as it doesn’t exceed 75 °C.
* The maximum power performance option is selected by the workstation.
* The simulation is being in single-precision and not double-precision.
* I would normally use 5000 for ntpr, ntwx and ntwr but increasing to 50000 did not make a difference.
* Utilising more CPU cores made performances worse.
Is there anything else that I have missed or could do to improve simulation performance.
Any suggestions are greatly appreciated 😊
Regards
Dan.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Nov 04 2025 - 08:30:02 PST
