You sure you are using the DPDP precision model? - You have to use that if
you want a direct comparison with the CPU code, otherwise you are probably
just seeing rounding differences. Here's what I have for CPU DPDP and SPDP:
This is for $AMBERHOME/test/cuda/chamber/dhfr_cmap
CPU
NSTEP = 0 TIME(PS) = 0.000 TEMP(K) = 359.72 PRESS =
0.0
Etot = -3135.7721 EKtot = 2230.2661 EPtot =
-5366.0381
BOND = 129.4500 ANGLE = 439.9122 DIHED =
754.0478
UB = 31.8773 IMP = 18.5859 CMAP =
-85.1050
1-4 NB = 367.0926 1-4 EEL = 6507.3376 VDWAALS =
-1009.1558
EELEC = -10036.4148 EGB = -2483.6659 RESTRAINT =
0.0000
DPDP
NSTEP = 0 TIME(PS) = 0.000 TEMP(K) = 359.72 PRESS =
0.0
Etot = -3135.7721 EKtot = 2230.2661 EPtot =
-5366.0381
BOND = 129.4500 ANGLE = 439.9122 DIHED =
754.0478
UB = 31.8773 IMP = 18.5859 CMAP =
-85.1050
1-4 NB = 367.0926 1-4 EEL = 6507.3376 VDWAALS =
-1009.1558
EELEC = -10036.4148 EGB = -2483.6659 RESTRAINT =
0.0000
These two match identically as one would expect.
SPDP
NSTEP = 0 TIME(PS) = 0.000 TEMP(K) = 359.72 PRESS =
0.0
Etot = -3135.7797 EKtot = 2230.2661 EPtot =
-5366.0458
BOND = 129.4500 ANGLE = 439.9122 DIHED =
754.0478
UB = 31.8773 IMP = 18.5859 CMAP =
-85.1050
1-4 NB = 367.0926 1-4 EEL = 6507.3376 VDWAALS =
-1009.1559
EELEC = -10036.4166 EGB = -2483.6715 RESTRAINT =
0.0000
SPDP has differences in VDWAALS in last decimal place, which is 8th sig fig
and EELEC in 3 decimal place which is also 8th sig fig. EGB differs in 2nd
decimal place which is 6th sig fig but this is expected since it is also a
function of the GB radii so there are more terms involved than the VDWAALS
or EELEC.
Can you try again and post the differences you see as well as your GPU specs
and compiler versions.
All the best
Ross
> -----Original Message-----
> From: Thomas Cheatham [mailto:tec3.utah.edu]
> Sent: Sunday, May 27, 2012 9:23 PM
> To: Ross Walker
> Subject: Re: [AMBER-Developers] [AMBER] manual prmtop file editing for
> free energy calculation
>
>
> dhfr_cmap
>
> ..had the differences...
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sun May 27 2012 - 23:00:02 PDT
