NVIDIA Shifts To Open-Source GPU Kernel Modules

NVIDIA fully moves to open source GPU core modules

NVIDIA has taken an important step in its commitment to open source software. The company announced that its upcoming R560 driver will be fully switched to open source GPU core modules . This move marks a significant shift in NVIDIA's strategy for driver development and distribution.

Table of contents

- background

• Progress and improvements
• Supported GPUs
• Installer changes
  • 1. Package Manager using CUDA Metapackage
  • 2. Run file installation
  • 3. Install Assistant Script
  • 4. Package Manager Details
  • 5. Windows Subsystem for Linux
  • 6. CUDA toolkit installation
• in conclusion

background

In May 2022, NVIDIA introduced an open source Linux GPU core module in the R515 driver. These modules are released under dual GPL and MIT licenses and were originally targeted at data center computing GPUs. At that time, support for GeForce and workstation GPUs was in the alpha stage.

Progress and improvements

Over the past two years, NVIDIA has made substantial progress:

• Performance : The performance of open source modules has now reached or exceeded the performance of closed source drivers.
• New features :
  • Heterogeneous memory management (HMM) support,
  • Confidential computing function,
  • Support for coherent memory architectures on the Grace platform.

Supported GPUs

The transition to open source modules has different effects on different GPU generations:

• Cutting-edge platforms : Grace Hopper and Blackwell platforms require open source modules.
• Supported GPUs : Newer architectures such as Turing, Ampere, Ada Lovelace and Hopper are fully supported by open source modules.
• Unsupported GPUs : Legacy GPUs from Maxwell, Pascal, and Volta architectures require continued use of proprietary drivers due to compatibility limitations.
• Hybrid deployment : Systems with a mix of old and new GPUs should continue to use proprietary drivers for optimal performance and stability.

If you are not sure which driver to install, don't worry! NVIDIA provides a detection assistant script to guide users to select the right driver.

Installer changes

NVIDIA is changing the default installation method for all installation methods from proprietary drivers to open source drivers.

1. Package Manager using CUDA Metapackage

When installing CUDA toolkits using the package manager, the top-level cuda package installs both the CUDA toolkit and the associated driver version. For example, installing cuda during CUDA version 12.5 provides the proprietary NVIDIA driver 555 and CUDA toolkit 12.5.

Previously, using open source GPU core modules required the installation of the distribution-specific NVIDIA driver open package and the selected cuda-toolkit-XY package.

Starting with CUDA 12.6, this process has changed. The default installation now includes open source drivers.

2. Run the file installation

The .run file installer for CUDA or NVIDIA drivers is now:

1. Query your hardware,
2. Automatically install the most suitable drivers.
3. Provides UI switching to choose between proprietary and open source drivers.

For command line or automated installations (such as Ansible ), use the following override:

 <code># 用于CUDA安装sh ./cuda_12.6.0_560.22_linux.run --override --kernel-module-type=proprietary # 用于NVIDIA驱动程序安装sh ./NVIDIA-Linux-x86_64-560.run --kernel-module-type=proprietary</code>

3. Install the Assistant Script

NVIDIA provides an assistant script to guide driver selection. To use it, first install the nvidia-driver-assistant package and then run the script:

 <code>$ nvidia-driver-assistant</code>

4. Package Manager Details

NVIDIA recommends using a package manager for consistent CUDA toolkit and driver installation. Here are the release-specific instructions:

Debian-based system :

Install open source drivers:

 <code>$ sudo apt-get install nvidia-open</code>

For Ubuntu 20.04, first upgrade to the open kernel module, and then install the open source driver like this:

 <code>$ sudo apt-get install -V nvidia-kernel-source-open $ sudo apt-get install nvidia-open</code>

RHEL-based system :

Install open source drivers:

 <code>$ sudo dnf module install nvidia-driver:open-dkms</code>

To upgrade using CUDA metapackage, disable module flow:

 <code>$ echo "module_hotfixes=1" | tee -a /etc/yum.repos.d/cuda*.repo $ sudo dnf install --allowerasing nvidia-open $ sudo dnf module reset nvidia-driver</code>

SUSE or OpenSUSE :

Select the appropriate command according to your kernel:

 <code># 默认内核版本$ sudo zypper install nvidia-open # Azure内核版本(sles15/x86_64) $ sudo zypper install nvidia-open-azure # 64kb内核版本(sles15/sbsa)适用于Grace-Hopper $ sudo zypper install nvidia-open-64k</code>

5. Windows Subsystem for Linux

WSL users do not need to do anything because it uses the NVIDIA kernel driver from the host Windows system.

6. CUDA toolkit installation

The installation process of the CUDA toolkit remains the same. Users can install it through their package manager as before.

 <code>$ sudo apt-get/dnf/zypper install cuda-toolkit</code>

For more detailed information about driver installation or CUDA toolkit settings, users can refer to the CUDA installation guide .

in conclusion

NVIDIA's move to open source GPU core modules marks a significant shift in the company's approach to driver development.

I really hope this will improve compatibility, performance, and user choice for a variety of GPU generations and Linux distributions.

resource :

• NVIDIA fully moves to open source GPU core modules

Featured image from Pixabay's Mizter_X94 .

The above is the detailed content of NVIDIA Shifts To Open-Source GPU Kernel Modules. For more information, please follow other related articles on the PHP Chinese website!