Using CUDA on AWS

Update: Includes additional CUDA packages that had been missed in original version.

At my new position with Replica Labs, we’re setting up some cloud-based systems for some heavy GPU computations and are leveraging the GPU instances provided by Amazon Web Services. For the most part, these are just headless Ubuntu 14.04 instances. I thought I’d document getting my process of getting proprietary NVIDIA drivers and the CUDA SDK installed and running here.

Disclaimer: These are the steps that I took, but they may not work for you. Information provided without warranty.

1. Use ubuntu-drivers to check what GPU model your server is equipped with.

Example output:

user@host:~$ sudo ubuntu-drivers devices
== /sys/devices/pci0000:00/0000:00:03.0 ==
vendor   : NVIDIA Corporation
model    : GK104GL [GRID K520]
modalias : ...
driver   : nvidia-331-updates - distro non-free
driver   : xserver-xorg-video-nouveau - distro free builtin
driver   : nvidia-331 - distro non-free recommended

So we can see that our AWS instance has an NVIDIA GRID K520 GPU.

1. Find the recommended driver version from NVIDIA at http://www.nvidia.com/Download/index.aspx?lang=en-us

2. Download and install the driver.
3. sudo sh NVIDIA-Linux-x86_64-331.89.run (insert your driver script name) to install the driver.
   • Don’t worry if the “preinstall” script fails, just continue.
   • Yes, install DKMS module
   • Yes, install 32-bit compatibility libraries
   • Yes, run nvidia-xconfig to configure the X server.
4. Install the CUDA packages: libcudart5.5, nvidia-cuda-dev and nvidia-cuda-toolkit.
5. Reboot the server.
6. Profit.

I have tried various other means of installing drivers including:

• Various of Ubuntu’s default repository drivers (331 and 331-updates)
• xorg-edgers PPA drivers (which said they were the correct version)

and none of them worked for me, unfortunately. I’d much rather install software through apt-get, but these drivers have forced me to do otherwise.