by chemal » 2014/01/08 16:17:32
* The cuda rpms from nvidia pile up everything under /usr/local/cuda where no rpm should install files. Plus they're not making the binaries and libraries available without fiddling with PATH and LD_LIBRARY_PATH. The run file gives you exactly the same in a clearer way.
* The cuda rpms from nvidia depend on nvidia's own rpm driver package, which is inferior to what elrepo has for el5/6 (to put it mildly).
Therefore the recommendations which I'm following myself:
1) Install nvidia driver from elrepo.
2) Install cuda from run file excluding the driver therein.
3) Use some method to adapt PATH and LD_LIBRARY_PATH (either in ~/.profile or for all users with a scriptlet in /etc/profile.d or use environment modules which allows you to even switch dynamically between different cuda versions if you have installed more than one).
Nvidia probably isn't mean, but their rpms are lousy.
Hope this helps.