NVIDIA drivers
AL2027 Preview
AL2027 is currently available for preview. It is intended for evaluation and testing only and is not recommended for production workloads.
Important
The AL2027 NVIDIA repositories do not yet support repository metadata signing.
Do not set repo_gpgcheck=1 for the NVIDIA repositories. If you do,
DNF can't load them and reports
repomd.xml GPG signature verification error: Bad PGP signature.
Packages in these repositories are still signed and verified by gpgcheck.
Signing support is planned for an upcoming release. For more information, see
Repository metadata signing in AL2027.
Amazon Linux 2027 provides NVIDIA GPU drivers and CUDA Toolkit packages through dedicated
repositories. AWS maintains these repositories and provides security advisories through
the Amazon Linux Security Center (ALAS)
Topics
About the NVIDIA repositories
For AL2027, we are moving from a single NVIDIA repository to a multi-repository structure:
-
One repository per NVIDIA GPU driver major version (
amazonlinux-nvidia-driver-). For examplemajoramazonlinux-nvidia-driver-580(Long Term Support Branch) andamazonlinux-nvidia-driver-595(Production Branch). Each repository follows the support timeline of the respective NVIDIA driver branch. You can enable repositories for different driver major versions at the same time. -
A products repository (
amazonlinux-nvidia-products) that provides everything except the GPU driver, such as CUDA Toolkit and related packages.
AWS qualifies this NVIDIA software with AL2027 release candidates before redistributing, and provides security advisories for the packages in these repositories.
New Feature Branch (NFB) drivers are skipped, because they do not
receive long-term security support. For more information about NVIDIA driver branches,
see NVIDIA
driver lifecycle
The repositories are available in all AWS Commercial Regions, as well as the AWS GovCloud (US) Regions and AWS China Regions.
The repositories provide NVIDIA Tesla (data center compute) and graphics drivers for
the x86_64 architecture only. GRID drivers, used for virtual display and
remote workstation capabilities, are not included. For GRID driver installation, see
NVIDIA
drivers for your Amazon EC2 instance in the Amazon EC2 User
Guide.
AL2027 Public Preview limitations
The following limitations apply to NVIDIA support on AL2027 during Public Preview:
Important
NVIDIA Fabric Manager (nvidia-fabricmanager), which coordinates
NVSwitch (the interconnect that links multiple GPUs on a single instance), is not
yet available in the AL2027 NVIDIA repositories. GPU instances that require
Fabric Manager (such as P4d, P5, and P6 instance types) are not yet supported on
AL2027.
NVIDIA Container Toolkit and NVIDIA Data Center GPU Manager (DCGM) are planned but are not yet available in the AL2027 NVIDIA repositories.
Choosing an NVIDIA driver version
Each NVIDIA GPU driver major version is served from its own repository. You enroll a
driver repository by installing an enrollment package. The enrollment package pulls in
the nvidia-release package as a dependency, which adds the products
repository (CUDA Toolkit and other NVIDIA products) and the NVIDIA GPG keys — so a
single install gives you both the driver and products repositories. Choose the
enrollment package that matches how you want to track driver versions:
| Enrollment package | Description |
|---|---|
nvidia-release-driver-prod |
Tracks the latest NVIDIA Production Branch driver major version. |
nvidia-release-driver-lts |
Tracks the latest NVIDIA Long Term Support Branch driver major version. |
nvidia-release-driver-latest |
Enrolls all supported driver major versions. The corresponding
repositories coexist at equal priority, so dnf installs the highest available
driver version and rolls forward across major versions as new ones are added. |
nvidia-release-driver- |
Enrolls a specific driver major version only (for example,
nvidia-release-driver-580). Use this to pin to a major version. |
To see which driver major versions are currently available, list the enrollment
packages. The numeric packages
(nvidia-release-driver-) correspond to
the driver major versions that AL2027 provides:
major
[ec2-user ~]$dnf list available "nvidia-release-driver-*"
To keep the AL2023 behavior — a single enrollment that makes all driver major
versions available, with dnf installing the latest and rolling forward as
new major versions are added — install nvidia-release-driver-latest:
[ec2-user ~]$sudo dnf install nvidia-release-driver-latest -y
For example, to track the latest Production Branch driver:
[ec2-user ~]$sudo dnf install nvidia-release-driver-prod -y
Note
When more than one driver-major repository is enabled, they coexist at equal
priority and dnf installs the highest available driver version
(Epoch-Version-Release), rolling forward to a newer major version on the next
dnf upgrade.
To pin a specific major version, enroll only that version
(nvidia-release-driver-) and do not
enroll majornvidia-release-driver-latest, -prod, or
-lts.
Removing nvidia-release-driver-latest does not by itself pin your
version. Because dnf distinguishes packages you install explicitly
from packages pulled in as dependencies, the outcome depends on how each major
version was enrolled:
-
A major version you installed explicitly stays enrolled after you remove
nvidia-release-driver-latest. If more than one remains enabled, the highest version still wins. -
A major version that was pulled in only as a dependency of
nvidia-release-driver-latestis removed along with it. If the one enrollment left is the major version you want, you are effectively pinned to it.
Installing NVIDIA drivers
After enrolling a driver repository, you can install NVIDIA driver packages using
dnf.
-
Install the kernel headers and development packages for your running kernel:
[ec2-user ~]$sudo dnf install kernel-devel-$(uname -r) kernel-headers-$(uname -r) -y -
Install the NVIDIA driver:
[ec2-user ~]$sudo dnf install nvidia-driver-cuda -y -
Reboot the instance:
[ec2-user ~]$sudo reboot -
After rebooting, verify the driver is loaded:
[ec2-user ~]$nvidia-smi
Note
To enable NVIDIA persistence mode, enable the nvidia-persistenced
service (provided by the driver repository):
[ec2-user ~]$sudo systemctl enable --now nvidia-persistenced
Installing CUDA Toolkit
CUDA Toolkit is served from the products repository. If you enrolled a driver
repository as described in Choosing an NVIDIA driver version, the products
repository is already present (nvidia-release was installed as a
dependency), so you can install CUDA Toolkit directly:
[ec2-user ~]$sudo dnf install cuda-toolkit -y
To verify that the products repository is available, run dnf repolist. You
should see the amazonlinux-nvidia-products repository in the list:
repo id repo name status
amazonlinux Amazon Linux 2027 repository enabled
amazonlinux-nvidia-products NVIDIA Products repository for Amazon Linux 2027 enabled
Removing the NVIDIA repositories
All NVIDIA repository packages — the products repository and each driver repository
enrollment packages — depend on nvidia-release. Removing
nvidia-release therefore removes every NVIDIA repository configuration
package in the same dnf transaction, whatever combination you enrolled:
[ec2-user ~]$sudo dnf remove nvidia-release -y
To remove only a single driver repository, remove that
driver's enrollment package instead. dnf also removes any alias enrollment
package (nvidia-release-driver-prod, -lts, or
-latest) that depends on the major version you remove:
[ec2-user ~]$sudo dnf remove nvidia-release-driver-major-y
Important
Removing repository configuration does not remove NVIDIA packages that are already installed, such as the GPU driver or CUDA Toolkit. Remove those packages separately if you no longer need them.