Installing CUDA on Ubuntu

Its simple as counting 1, 2, 3.. (as long as everything is supported and works fine).

PART I:
NVidia (developer) driver that is required by CUDA (as mentioned in the GPU computing page). This might be different from the latest driver you might find in other pages.

Step 1: Download the driver for your nvidia graphics card. These will come as .run files.

Step 2: If you have traces of your (now obsolete) driver from nVidia, get rid of them with the following 2 commands:
sudo apt-get remove --purge nvidia
sudo nvidia-installer --uninstall

Step 3: Stop your display manager
sudo /etc/init.d/gdm stop

Step 3.5: Now you do not have your X running, so switch to console:
ctrl + alt + F1

Step 4:
sudo sh devdriver_3.2_linux_32_260.24.run

This should bring up a console based miniature GUI, enter details as required. And complete the installation.

Step 5: Bring back the X
sudo /etc/init.d/gdm start


PART II: Installing CUDA toolkit
Step 1: Download the CUDA toolkit (cudatoolkit_3.2.9_linux_32_ubuntu10.04.run) from GPU Computing page at NVidia site.

Step 2: Run the file:
sh cudatoolkit_3.2.9_linux_32_ubuntu10.04.run

Make sure you do not do it as root (or with sudo prefix).
Step 3: I prefer leaving the default locations for installation.

Step 4: Assuming default locations of installation, set environment variables

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib


PART III: Installing sample projects, SDK etc.

Step 1: Download gpucomputingsdk_3.2_linux.run

Step 2: Run (do not run this in root mode)
sh gpucomputingsdk_3.2_linux.run

Step 3: Give default locations to CUDA installation (or the modified locations, if you have modified them)

Step 4: Test if its fine
cd ~/NVIDIA_GPU_Computing_SDK/C
make

This should build all example projects included in the SDK without any error. If that goes fine, it means you have a working CUDA environment to work with.

Now try running one of the programs (one I like is to query the GPU)
cd bin/linux/release/
./deviceQuery

It gives me an output like this:
./deviceQuery Starting...

CUDA Device Query (Runtime API) version (CUDART static linking)

There is 1 device supporting CUDA

Device 0: "GeForce GT 240M"
CUDA Driver Version: 3.20
CUDA Runtime Version: 3.20
. . . . .
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 3.20, CUDA Runtime Version = 3.20, NumDevs = 1, Device = GeForce GT 240M



Components that I have/chose:

1. Ubtuntu 10.04 (32bit)
   
2. NVidia GT240M graphics card
3. devdriver_3.2_linux_32_260.24.run driver
4. cudatoolkit_3.2.9_linux_32_ubuntu10.04.run
5. gpucomputingsdk_3.2_linux.run
   

Read more...