We will likely need to follow the same procedure as we did for the Windows Nvidia driver bots in  issue 771654 :

1) specialize the GPU bot json on driver version, and add an experimental version
2) add a new Nvidia experimental bot on the Linux GPU FYI waterfall.
3) perform a quiesced upgrade a handful of machines
4) update the GPU bot json for the experimental server to the new driver version
5) once stable, quiesce and upgrade half the fleet to the new version
6) update the GPU bot json for the stable version to the new driver version
7) quiesce and upgrade the second half of the fleet

Feel free to steal this Ken if you plan to work on it.