A big-bytes thank-you to SuperMicro for having delivered a GPU system for testing. The system delivered is a SuperMicro – Tesla M1060. OS choice was Scientific Linux 5.4 x86_64. The distribution choice was chosen based on our existing HPC and GRID kit.
Installation of the drivers and various kits was very simple and documented below.
Firstly, download the Nvidia GPU drivers from Nvidia Driver WebSite. You would then need to execute “ chmod +x NVIDIA-Linux-x86_64-280.13.run “. Execute the binary “ ./NVIDIA-Linux-x86_64-280.13.run “. The installation will complain that the kernel headers have not been installed. The installation will terminate and now allow for you to install the kernel headers. Once the kernel headers are installed re-run the installation. You may verify the install by executing the system management interface tool, nvidia-smi.
“nvidia-smi -L “ will list all available GPU's configured on the system.
[root@host installs]# nvidia-smi -L
GPU 0: Tesla M1060 (UUID: GPU-80671de8eb0d438c-5f7dde2b-2ea53740-1d3b870a-86bf5cc98531e86374bc0514)
GPU 1: Tesla M1060 (UUID: GPU-d4df9bac30037980-47cc692d-848f2bec-851c259f-fbfbd47d360bd2021f066d91)
Next, we install the Cuda Toolkit. As of this post its on version 4.0. Another simple install. Download it from http://developer.nvidia.com/cuda-toolkit-40#Linux. Again “ chmod +x cudatoolkit_4.0.17_linux_64_rhel6.0.run “ and then “./cudatoolkit_4.0.17_linux_64_rhel6.0.run” . The default install location is /usr/local/cuda. Added the following lines below to the /etc/ld.so.conf file and execute, as root, “ldconfig” to update ld information.
Finally, we install the Cuda SDK found on the same web page as the Cuda Toolkit. The instructions for this is exactly the same aside from the filename.
A side note: The SMI tool may return many N/A entries and this is a result of the latest Nvidia driver. We will provide a series of results from our GPU testing in the coming weeks.