by Watchanan Chantapakul
Hi folks!
To make the CUDA works on a computer might be a pain for someone including me, so I just want to memorise the way I make the system work without any problem. In this post, it is based on these requirements:
The first thing you need to do is to check the compability between Cuda Toolkit version and Linux x86_64 driver version. You can quickly look up the table here. In the case that you want to look at other details, you can go to CUDA Compatibility :: GPU Deployment and Management Documentation.
For instance, now (2018-12-18), the most up-to-date CUDA version is 10.1, so I will stick to this version, then I check the aforementioned table. As a result, I have to install Nvidia driver for Linux 410, as It says (CUDA 10.0 (10.0.130) >= 410.48).
To install the Nvidia driver 410, just type this (debian-based):
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt install nvidia-driver-410
And do not forget to reboot your Linux with sudo reboot command.
Firstly, you have to download the CUDA Toolkit installer. Note that the version of CUDA Toolkit must be matched with your Nvidia driver version, and vice versa. There are about 4 different ways to install it, so you can easily follow its instruction, but the one that I recommend is to use runfile (local) approach as there are only 2 steps to complete it.
cp /usr/local/cuda/cuda-testing/samples
make
You can make the make comamnd run faster by adding a flag -j?. The ? is the number of parallel threads you want it to operate, so the more threads, the more processors you assign it to use. For instance, if my CPU has 8 cores, I can run make command as make -j8.
Let’s run an example to check whether it works flawlessly with this command:
./bin/x86_64/linux/release/vectorAdd