Good day, I tried to install CUDA 12.6 on Linux Mint 2022 and encountered this problem. After command:
cmake .. -DGMX_GPU=CUDA -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda
I got an output:
CMake Error at /usr/share/cmake-3.28/Modules/CMakeDetermineCUDACompiler.cmake:270 (message):
Failed to detect a default CUDA architecture.
Compiler output:
Call Stack (most recent call first):
cmake/gmxManageCuda.cmake:116 (enable_language)
CMakeLists.txt:697 (include)
What could be the problem here? I have selected the version of the download file CUDO as in the picture.
Thank you
Hi Roman Shaykenov
First of all, do you have a CUDA compatible GPU?
Hello Nicola Piasentin
My GPU is Nvidia Geforce 117M. I don`t know, is it compatible with CUDA.
I think most - if not all - of the NVIDIA GPUs are compatible with CUDA.
Even before going through your problem, by having a quick glance online the GF117 GPU seems to be a Fermi, Kepler, or Maxwell architecture, I am a bit unsure, but most likely a Fermi one. At the 'GPU Supported' table in this wiki page
https://en.wikipedia.org/wiki/CUDA
you can see that Fermi compatibility stops at 8.0, while in the image you posted you are downloading the 12.6. If you really have a Fermi architecture (the next table in the same page would say so, as the GF117 is listed under Fermi), you can only have CUDA SDK version up to 8.0.
So, if you really need CUDA, I would go for the 8.0 version. A few key points, though. Given the very old GPU, if you have a modern Ubuntu (22/24) you could have problems with compilers to install such an old version, but eventually it might be doable. Then, I am quite sure GROMACS does not support such old CUDAs (e.g. for GRO24.X min version is CUDA 11.0), and it is likely that you will also have to substantially downgrade GROMACS to an older version that supports CUDA < 9.0, which will also have several problems, like finding proper old compilators, old installation guides, and - most important- renounce all the improvement in speed and stability of the newer GROMACS versions. Lastly, even if you manage to go around all of this, I am unsure about any gain in speed that you could get from such an old architecture. I mean, it will be most likely faster than pure CPU, but still the price to pay in all other terms seems a lot to me.
So if I were you I would double check the GPU version and the CUDA compatibility, and check which is the newest GROMACS version you can get wrt these parameters, as forcing the GPU compatibility may not be worth the prize!
The error suggests that GROMACS cannot automatically detect your CUDA architecture. To solve this:
all of which can we easily done at mdsim360.com, a new platform that lets you run MD simulations entirely online without local installation.
- Badges
- Science method