On 04/18/2018 04:21 AM, Bjoern Forsberg wrote:
> Just to elaborate on Markus solution, I would also STRONGLY recommend
> sticking to supported configs
> <https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>.
> It's a bit annoying that NVIDIA hasn't made support for later
> GCC-versions more of a priority, *but the version-check is there for a
> reason*.
Installing an older compiler gets around the compiler version check, but
I wonder what impact glibc has on this; as that's harder to change.
Ubuntu 18.04 ships with glibc v. 2.27, and according to the supported
configs URL posted above, Nvidia only supports up to 2.24-3.
The Ubuntu thing is a bit of an annoyance right now, as Ubuntu 16.04
includes a couple of major abandoned technologies: Unity and Upstart,
which I would prefer to avoid, not the least because they're both tech
junk that should have never made it off the drawing board.
Also, we recently purchased some new processing workstations with Titan
X cards (the only consumer grade card we could get), and I have been
unable to get 16.04 to provide a stable user interface on these systems
with the current Nvidia drivers; hence had to switch to 18.04.
If it were up to me, everyone in the scientific computing world would be
running Arch linux, with applications which can't be recompiled running
in containers. Way fewer wasted brain cycles in the long run, and if
Nvidia can't keep up, recode everything from CUDA to OpenCL and open the
door for other GPU vendors. I suspect Nvidia's game would improve
considerably and quickly as a result.
|