GPU error during installation of Gromacs 2020.3

User discussions
1

GROMACS version:2020.3
GROMACS modification: Yes/No
Here post your question
Hi guys, I’m having trouble installing GPU enabled gromacs on my computer (running ubuntu 20.04, GPU: NVIDIA GeForce RTX 2080). I got an error:
nvcc fatal : Unsupported gpu architecture 'compute_30'
Seems like the error is because Cuda (version 11) doesn’t support ‘compute_30’. After some searches, I tried passing following flags in CMake
-DGMX_CUDA_TARGET_COMPUTE=50 as well as -DGMX_CUDA_TARGET_COMPUTE=60
However, none of them worked: the installation failed in one of the tests
12 - GpuUtilsUnitTests (Failed)
I’m wondering if someone has experienced a similar issue and might have any solution?

2

You’re right that it’s likely an incompatibility of Gromacs with CUDA 11 - it just came out, after all. I suggest downgrading to CUDA 10 if you can. I have access to a 2080 system with that OS so I’ll see if I can get it working and report back here if it’s an easy patch, but otherwise you might have to wait until the next Gromacs release.

1 Like
3

edit “gmxManageNvccConfig.cmake”

as like this

# Set the CUDA GPU architectures to compile for:
# - with CUDA >=9.0         CC 7.0 is supported and CC 2.0 is no longer supported
#     => compile sm_61, sm_70 SASS, and compute_70 PTX
# - with CUDA >=10.0        CC 7.5 is supported
#     => compile sm_61, sm_70, sm_75 SASS, and compute_75 PTX

# First add flags that trigger SASS (binary) code generation for physical arch
list (APPEND GMX_CUDA_NVCC_GENCODE_FLAGS "-gencode;arch=compute_61,code=sm_61")
list (APPEND GMX_CUDA_NVCC_GENCODE_FLAGS "-gencode;arch=compute_70,code=sm_70")

# Next add flags that trigger PTX code generation for the newest supported virtual arch
# that's useful to JIT to future architectures
list (APPEND GMX_CUDA_NVCC_GENCODE_FLAGS "-gencode;arch=compute_61,code=compute_61")
list (APPEND GMX_CUDA_NVCC_GENCODE_FLAGS "-gencode;arch=compute_70,code=compute_70")
if(NOT CUDA_VERSION VERSION_LESS "10.0")
    list (APPEND GMX_CUDA_NVCC_GENCODE_FLAGS "-gencode;arch=compute_75,code=compute_75")
endif()

It’s gonna force gromacs to use compute 6.1 or u can use compute 7.0 for RTX 2080

1 Like
4

The important part there is removing the sm_30 compilation flag from both sections, since it’s no longer supported in CUDA 11. Removing sm_35,37,50,52 is optional - it’ll give you a deprecation warning if you compile with it, but shouldn’t crash

1 Like
5

Thank you both for your responses. I edited the gmxManageNvccConfig.cmake file as suggested. But the installation still failed the Test #12: GpuUtilsUnitTests (3 failed, 70 passed) :

[ FAILED ] 3 tests, listed below:
[ FAILED ] HostAllocatorTestCopyable/0.ManualPinningOperationsWorkWithCuda, where TypeParam = int
[ FAILED ] HostAllocatorTestCopyable/1.ManualPinningOperationsWorkWithCuda, where TypeParam = float
[ FAILED ] HostAllocatorTestCopyable/2.ManualPinningOperationsWorkWithCuda, where TypeParam = gmx::BasicVector

Errors while running CTest
make[3]: *** [CMakeFiles/run-ctest-nophys.dir/build.make:58: CMakeFiles/run-ctest-nophys] Error 8
make[2]: *** [CMakeFiles/Makefile2:2486: CMakeFiles/run-ctest-nophys.dir/all] Error 2
make[1]: *** [CMakeFiles/Makefile2:2465: CMakeFiles/check.dir/rule] Error 2
make: *** [Makefile:249: check] Error 2

I went ahead and installed the gromacs anyway but I got segmentation fault error during runtime.

Subin

6

CUDA 11 is now supported from the head of github master but isn’t available in a stable release yet. Do you have access to CUDA 10? It should be fine to have the CUDA 11 drivers on the device but build against CUDA 10.

1 Like
7

I downgraded to CUDA 10 and installed gromacs2020.3. However, there was an issue with the ‘renumbering of residue number to 1’. So I installed gromacs2019.6 instead and have been using it since. I will wait until new version comes out to upgrade. Thanks!

8

Hi All,

I am facing the same problem with Gromacs 2020.3 and CUDA 11 on Ubuntu 20.04 LTS.
Is there any workaround available yet?

Thanks.

Raman

9

please , did you solve it ?

10

No more workarounds are needed, just get the latest release (2020.4).

1 Like
11

thanks, i will try it now … but does it will be compitabele with nvidia rtx 2060 cuda 11.1 ?

12

finally, it is solved … Many thanks

13

I used GCC 7 and CUDA 10.2. The newer version of Gromacs 2020.4 seems to be CUDA 11 compatible but I didn’t give it a try yet.

14

Yes, the CUDA 11 compatibility issues have been addressed in the 2020.4 release.

15

The code was running with gpu and suddenly screen of ubuntu gave me log in … I tried to log in but couldn’t until I restart laptop and got
Starting authorization manager

I know it’s a problem of ubuntu … but I had two graphic card one is nvidia and another is intel … does that because of enforce gpu for working… can you help me because I was happy that code is running and afraid of deleting or doing something with make me stuck again

Appreciate you help