Dear all,
I have successfully build the Amber22 pmemd.cuda code on a system with NVIDIA RTX4090.
It indeed runs extremely fast. However, I noticed that it produces output quite different from results using CPU pmemd or using a RTX3080 card.
This is the output for one MD step on a protein system using pmemd.cuda on RTX4090 (built with cuda11.7, gcc 11.3 ubuntu22.04):
NSTEP = 1 TIME(PS) = 602900.002 TEMP(K) = 348.44 PRESS = 26341.6
Etot = -150709.4789 EKtot = 31938.7695 EPtot = -182648.2485
BOND = 599.7020 ANGLE = 1279.5888 DIHED = 1332.9130
UB = 0.0000 IMP = 0.0000 CMAP = 44.8633
1-4 NB = 524.8852 1-4 EEL = -715.5685 VDWAALS = 20678.9140
EELEC = -206393.5464 EHBOND = 0.0000 RESTRAINT = 0.0000
EKCMT = 15008.1735 VIRIAL = -247556.8301 VOLUME = 461654.7275
Density = 1.0066
------------------------------------------------------------------------------
using pmemd (CPU) or pmemd.cuda built on a machine with RTX3080 gives (for same input and restart):
NSTEP = 1 TIME(PS) = 602900.002 TEMP(K) = 339.68 PRESS = -117.2
Etot = -150385.9274 EKtot = 31136.6108 EPtot = -181522.5382
BOND = 599.7020 ANGLE = 1279.5888 DIHED = 1332.9127
UB = 0.0000 IMP = 0.0000 CMAP = 44.8633
1-4 NB = 524.8852 1-4 EEL = -715.5685 VDWAALS = 20678.9428
EELEC = -205267.8647 EHBOND = 0.0000 RESTRAINT = 0.0000
EKCMT = 15008.1735 VIRIAL = 16176.1626 VOLUME = 461654.7275
Density = 1.0066
The main difference is the pressure, however, even constant volume runs give very different results after a few MD steps. I also tried to build pmemd.cuda using ubuntu 20.04, cuda11.5 and gcc 9.4 (on RTX4090 system, I use the nvidia-525 driver). I get the same output as for cuda11.7, gcc 11.3 ubuntu22.04.
I probably do something wrong but I don't know what?
Thanks
Martin Zacharias
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Feb 23 2023 - 02:00:03 PST
