I’m using nvml to support computing environment for our users and planning to offer CUDA version information to them by implemented software including NVML.
However, I just can get the version information about GPUs driver and nvml only by NVIDIA Management Library function, so I’m not quite sure that I can give the right version information to my users.
As I understand, since CUDA toolkit contains NVIDIA Management Library, if I install new CUDA version on my environment, the NVML version would be same as the CUDA version but I concern … a possibility that it will be varied each other.
CUDA toolkit contains only header file “include/nvml.h” and “stub” library “lib64/stubs/libnvidia-ml.so” (null functions, check with “readelf -a lib64/stubs/libnvidia-ml.so”).
[root@centos75 /usr/local/cuda]# ls -l lib64/stubs/libnvidia-ml.so
-rwxr-xr-x. 1 root root 31664 Oct 28 21:22 lib64/stubs/libnvidia-ml.so
True library is distributed ONLY with driver.
[root@centos75 /usr/local/cuda]# ls -l /usr/lib64/libnvidia-ml.so*
lrwxrwxrwx. 1 root root 17 Mar 18 17:26 /usr/lib64/libnvidia-ml.so -> libnvidia-ml.so.1
lrwxrwxrwx. 1 root root 23 Mar 18 17:26 /usr/lib64/libnvidia-ml.so.1 -> libnvidia-ml.so.390.115
-rwxr-xr-x. 1 root root 1314560 Mar 18 17:26 /usr/lib64/libnvidia-ml.so.390.115
There can be only ONE driver and ONE NVML library installed (driver from standard distribution or driver from CUDA package (optional)).