This is more to confirm something rather than a question - the last time I built an RPM was years ago and usually it was noarch.

Now we're talking about NIC drivers. So the best is that compatibility is respected. Mellanox provides a tarball with an installer that compiles the module in your server.

The nice part is that the installer creates RPMs, which it uses later :

/root/nic/mlnx-en-2.1-1.0.0/RPMS/rhel6u3/x86_64/mlnx_en-2.1-1.0.0.0.gfeee0c2.2.6.32_431.17.1.el6.x86_64.x86_64.rpm
/root/nic/mlnx-en-2.1-1.0.0/RPMS/rhel6u3/x86_64/mlnx_en-devel-2.1-1.0.0.0.gfeee0c2.2.6.32_431.17.1.el6.x86_64.x86_64.rpm

I imagined that instead of recompiling that in every server reinstall, I could pretty much add these 2 RPMs to a YUM repository and have it installed in a more automated way, my scenario :

* All servers with the same "OS"(distro/release) - GCC and Kernel will be always the same
* As I'm using CentOS generic kernel it won't matter but for the record - all servers use the same processor.

I think that all is fine, but I wanted to confirm with other people that I'm not overlooking any problem. If you do see one, let me know, thanks!

Given the fact servers usually need a reboot to run a new kernel personally I'd go the kmod route, otherwise there indeed is nothing with your scenario (except its totally static and you'll need to patch, rebuild, test, package the .ko anyway if you decide to install a new kernel version).