nvcc fatal: Value 'sm_61' is not defined for option 'gpu-architecture'

[carmine marco Morgillo]

Hi,
I have installed soMD and FESetup just started with the tutorial on running MD.
When I recall the software to run the nvt equilibration I receive this error:
RuntimeError: Error launching CUDA compiler: 256
nvcc fatal : Value 'sm_61' is not defined for option 'gpu-architecture'

This is the details of my cuda path:
[acellera@localhost Equil]$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2015 NVIDIA Corporation
Built on Tue_Aug_11_14:27:32_CDT_2015
Cuda compilation tools, release 7.5, V7.5.17

[acellera@localhost Equil]$ sudo find / -name nvcc
[sudo] password for acellera:
/usr/local/cuda-7.5/bin/nvcc

[acellera@localhost Equil]$ echo $PATH
/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/opt/acellera/miniconda3/bin:/usr/lib64/openmpi/bin:/home/acellera/sire.app/bin:/usr/local/cuda-7.5/bin

[acellera@localhost Equil]$ echo $LD_LIBRARY_PATH
:/usr/local/cuda/lib64:/usr/local/cuda/lib64


OpenMM seems to recognize the CUDA:
In [1]: import simtk.openmm as mm
In [2]: platforms = [ mm.Platform.getPlatform(index).getName() for index in range(mm.Pl
...: atform.getNumPlatforms()) ]
In [3]: print (platforms)
['Reference', 'CPU', 'CUDA', 'OpenCL']

The GPU cluster is equipped with 4xGTX1080 and the OS is CentOS7.4, the driver installed are CUDA7.5 and nvidia driver version 384.81.

[julie...@googlemail.com]

Hi Marco, 

Can you give more details about how you installed Sire ? A mismatch between CUDA library versions used to build SOMD and those available on the GPU nodes is often the cause of this problem. Try testing the CPU platform  to see if it works (will be slow but will help troubleshoot). 

Specify also what tutorial you are following, some of the online tutorials need to be refreshed and may not be consistent with the current API. 

Best wishes,

Julien

[carmine marco Morgillo]

Hi Julien,

I downloaded the auto-installer from the sire page https://siremol.org/pages/download.html following the first choice "Download a sire binary". The installation reported no problems.
Hence I followed the tutorial described here https://siremol.org/tutorials/somd/MD/Equib.html but I was stuck at nvt equilibration part.

Editing the option "platform = GPU" in "platform = CPU" gives the same error as before.
This is the verbose output
[acellera@localhost Equil]$ somd -C cpu_nvt.cfg
Starting somd: number of threads equals 16

Loading configuration information from file cpu_nvt.cfg
Using parameters:
===============
# energy frequency == 100
#Snapshots == 5000/100 = 500 snapshots.
andersen == True
andersen frequency == 10
barostat == False
barostat frequency == 25
buffered coordinates frequency == 100
center of mass frequency == 10
constraint == hbonds
crdfile == 1AKI.rst7
cutoff distance == 10 angstrom
cutoff type == cutoffperiodic
integrator type == leapfrogverlet
minimise == False
ncycles == 10
nmoves == 500
platform == CPU
pressure == 1 atm
reaction field dielectric == 78.3
save coordinates == True
temperature == 298.15 t
timestep == 0.001 ps
topfile == ../1AKI.parm7
use_restraints == True
===============

### Running Molecular Dynamics simulation on localhost.localdomain ###
###================Setting up calculation=====================###

==============================================================
Sending anonymous Sire usage statistics to http://siremol.org.
Index GPU = 0
For more information, see http://siremol.org/analytics
Loaded a restart file on which we have performed 0 moves.
To disable, set the environment variable 'SIRE_DONT_PHONEHOME' to 1
To see the information sent, set the environment variable
SIRE_VERBOSE_PHONEHOME equal to 1. To silence this message, set
the environment variable SIRE_SILENT_PHONEHOME to 1.
==============================================================

There are 25196 atoms in the group
###===========================================================###

###=======================somd run============================###
Starting somd run...
500 moves 10 cycles, 5 ps simulation time

Cycle = 1
Traceback (most recent call last):
File "/home/acellera/sire.app/pkgs/sire-2018.2.0/share/Sire/scripts/somd.py", line 125, in <module>
OpenMMMD.run(params)
File "/home/acellera/sire.app/lib/python3.5/site-packages/Sire/Tools/__init__.py", line 172, in inner
retval = func()
File "/home/acellera/sire.app/lib/python3.5/site-packages/Sire/Tools/OpenMMMD.py", line 1473, in run
system = moves.move(system, nmoves.val, True)

RuntimeError: Error launching CUDA compiler: 256
nvcc fatal : Value 'sm_61' is not defined for option 'gpu-architecture'

Thanks for your help,

Carmine