CUDA¶

CUDA is a mechanism for allowing general purpose computing on graphics cards rather than the normal CPU. It is thus optimised for parallel computing and high memory throughput. CUDA requires a CUDA-enabled graphics processing unit (GPU) which means a NVIDIA graphics card is required.

Installing CUDA¶

The NVIDIA CUDA software repository is available on the Linux desktop platform so you can install the various versions like this:

8.0: sudo yum install cuda-8-0
7.5: sudo yum install cuda-7.5
7.0: sudo yum install cuda-7-0

This installs the official NVIDIA CUDA packages. CUDA is installed into:

/usr/local/cuda-N.N/

Where N.N is the version, e.g. /usr/local/cuda-8.0. To add CUDA to your path add this to your .bashrc:

export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}}

Using CUDA Samples¶

Once installed you can deploy and build the CUDA samples like this:

mkdir ~/cuda
/usr/local/cuda-8.0/bin/cuda-install-samples-8.0.sh ~/cuda
cd ~/cuda/NVIDIA_CUDA*
make

That will build all the samples, which you probably don’t want. Instead change into a sample sub directory and build just one at at time, like so:

cd ~/cuda/NVIDIA_CUDA*/0_Simple/clock/
make

You can then run the compiled clock like so:

[db2z07@uos-212247 clock]$ ./clock
CUDA Clock sample
GPU Device 0: "GeForce GTX 750 Ti" with compute capability 5.0

Average clocks/block = 4042.890625