Re: [AMBER] Amber11: CUDA-enabled version does not work

From: Jason Swails <>
Date: Thu, 15 Jan 2015 07:11:49 -0500

On Thu, 2015-01-15 at 17:24 +0900, Khuong Truong Gia wrote:
> Hi David,
> This is result when I run "":
> ./
> Using default GPU_ID = -1
> Using default PREC_MODEL = SPDP
> make[1]: Entering directory `/opt/amber11/test'
> cd cuda && make -k test.pmemd.cuda GPU_ID=-1 PREC_MODEL=SPDP
> make[2]: Entering directory `/opt/amber11/test/cuda'
> ------------------------------------
> Running CUDA Implicit solvent tests.
> Precision Model = SPDP
> GPU_ID = -1
> ------------------------------------
> cd trpcage/ && ./Run_md_trpcage -1 SPDP netcdf.mod
> diffing trpcage_md.out.GPU_SPDP with trpcage_md.out
> ==============================================================
> cd myoglobin/ && ./Run_md_myoglobin -1 SPDP netcdf.mod
> diffing myoglobin_md.out.GPU_SPDP with myoglobin_md.out
> *.....(all Implicit solvent tests PASSED),Then, it stood still (no update
> for long time) at line "cd 4096wat/ && ./Run.pure_wat -1 SPDP netcdf.mod"*
> cd gb_ala3/ && ./Run.irest1_ntt1_igb1_ntc2 -1 SPDP netcdf.mod
> diffing irest1_ntt1_igb1_ntc2.out.GPU_SPDP with irest1_ntt1_igb1_ntc2.out
> ==============================================================
> ------------------------------------
> Running CUDA Explicit solvent tests.
> Precision Model = SPDP
> GPU_ID = -1
> ------------------------------------
> cd 4096wat/ && ./Run.pure_wat -1 SPDP netcdf.mod
> Is that my installation's mistake?

It is hard to tell. Your symptoms look a lot like problems that were
addressed by bugfix.12 for Amber 11 (see If you have not already applied all
of the bug fixes, you need to. If you have and it still doesn't work,
you probably need to upgrade to Amber 14 to use pmemd.cuda.

As Dave pointed out, Amber 11 is 5 years old at this point and
represents pmemd.cuda in its infancy. Amber 11 was the first release
that featured a production-ready GPU code. Support for that version
ended not too long after Amber 12 was released (almost 3 years ago now).

Since that time, GPU hardware has changed dramatically (Amber 11 was
released alongside the Tesla architecture, like the Tesla 1060 and 1070
cards and the GTX 2xx series). NVidia has gone through several
different generations of architectures (and CUDA toolkits) since then --
through Fermi, Kepler, and now on to Maxwell. Each architecture
necessitated changes to pmemd.cuda's internals in order to take better
advantage of the new hardware. As a result, Amber 11 won't run (or
won't run well) on Kepler or Maxwell cards. Furthermore, numerous bugs
have been found and fixed since Amber 11 has been released, and those
fixes will most certainly not be ported back to Amber 11. Performance
has also improved *dramatically* since the days of Amber 11.

Long story short, it is probably time to retire Amber 11 if you are
having problems with it. You are unlikely to find anybody on this list
that can help out that has used Amber 11 within the last 3 to 5 years. I
don't think Amber 11 is something we actively support anymore.

Hope this helps,

Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Thu Jan 15 2015 - 04:30:02 PST
Custom Search