From: Tristan Croll (tristan.croll_at_qut.edu.au)
Date: Tue Jan 13 2015 - 18:29:59 CST

Hi all,

Having just spent a very solid day banging my head against a wall to get a new laptop up and running with CUDA, I thought I'd share my experience here. CentOS 7 was a wash for me - while it may be possible, there will be a lot of manual adjustment of configuration files etc. required to get it going. For me (installing on an MSI GT60 2QD, with the Maxwell generation GTX970), following the instructions at http://blog.mdda.net/oss/2014/12/17/fedora21-from-scratch/ from a clean Fedora 21 install got most of the way there with little fuss. The only addition I had to make was to install the 343.16 NVidia driver downloaded from the NVidia website (this shouldn't be necessary for Kepler or earlier GPUs). Note that if using these, it's vitally important to include the --no-opengl-files switch and to choose NOT to update xorg.conf when prompted, otherwise you'll bork your desktop.

If you've followed the instructions correctly up to and including the "Make the Suspend of the NVidia GPU work" section (note that there's one missing step - you need to add your user to the "bumblebee" group), then you'll find that you can run CUDA builds of NAMD or VMD by prefixing with optirun, e.g.

optirun vmd

or

optirun namd2 <args>

This gives the program exclusive access to the NVidia GPU driver - as far as the rest of the system is concerned, the NVidia GPU doesn't exist.

The reward is a portable system that runs really quite surprisingly fast simulations, plus all the newest vmd goodies.

Cheers,

Tristan