Jan 2nd, 2022
- System Specification Check
- NVIDIA Driver Installation
- CUDA Toolkit Installation
- cuDNN Installation
- Nvidia Documentation
- Check your system architecture to select correct installers for your platform
$ uname -m $ dpkg --print-architecture
- Remove old installation
$ sudo apt-get purge nvidia-* $ sudo apt-get update $ sudo apt-get autoremove # DO NOT skip this line
- Search for latest version of Nvidia driver
$ apt search nvidia-driver
- Install Nvidia libraries
$ sudo apt install libnvidia-common-<version> $ sudo apt install libnividia-gl-<version>
- Install Nvidia driver
$ sudo apt install nvidia-driver-<version>
- Reboot and check for the installation
$ nvidia-smi
- Intsall kernel headers and developement packages for your currently running kernel
$ sudo apt-get install linux-headers-$(uname -r)
- Download and install CUDA Toolkit
- CUDA Toolkit from Nvidia Developer
- Select target platform
- Recommendation: pick deb [network] option of Installer Type
- Follow the installation instruction on the download page to install CUDA Toolkit
- To include GDS package with CUDA Toolkit
- GDS Overview: CUDA GPUDirect Storage
$ sudo apt-get install nvidia-gds
- Setup environment
- Config
$PATH
variable with following script:CUDA_HOME=/usr/local/cuda PATH=${CUDA_HOME}/bin${PATH:+:${PATH}} LD_LIBRARY_PATH=${CUDA_HOME}/lib64 ${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}} export LD_LIBRARY_PATH export CUDA_HOME export PATH
- Add the script to either:
~/.bashrc
for user session usage/etc/profile
for system wide usage
- Setup POWER9
- Check NVIDIA Persistence Daemon
$ systemctl status nvidia-persistenced
- If it is not loaded
$ sudo systemctl enable nvidia-persistenced
- Reboot and check for installation
$ nvcc --version
- Download cuDNN:
- Nvidia cuDNN from Nvidia Developer (local installer)
- NVIDIA Developer Program Membership is required to download
- Select CUDA matching version and target platform
- Install cuDNN
- Import CUDA GPG key
$ sudo dpkg -i <downloaded-file> $ sudo apt-key add /var/cudnn-local-repo-*/7fa2af80.pub $ sudo apt-get update
- To auto-match version of cuDNN v8 with version of CUDA when installing:
$ sudo apt-get install libcudnn8 $ sudo apt-get install libcudnn8-dev $ sudo apt-get install libcudnn8-samples
- Operating System: Ubuntu 20.04 x84_64 (64-bit)
- Architecture: amd64
- GPU: Nvidia GeForce GTX 1050
- Installation with success on: Jan 2nd, 2022
Btw, did you need to enable "compatibility" mode on 3090, or was it working without it too?
As per docs, with "compatibility" mode enabled GDS's IO path will fall back to the traditional CPU path..