Alauda GPU Management

Compatibility

Use this page to check whether an Alauda GPU Management deployment matches the supported product scope.

TOC

Product support scopeValidation tiersPrecompiled driver image compatibilityKnown limits

Product support scope

ItemSupported scope
ACPv4.0-v4.3
Architectureamd64 (x86_64)
Driver modePrecompiled driver (immutable OS), preinstalled host driver, or upstream driver (mutable OS)
Runtime integrationDevice plugin allocation of nvidia.com/gpu with the nvidia RuntimeClass configured by the container toolkit
MonitoringDCGM-Exporter; ACP owns the GPU dashboards
Node discoveryAlauda Build of Node Feature Discovery as a separate prerequisite plugin (not in this package); GPU Feature Discovery is delivered by the package
Optional componentsMIG (user-supplied mig-manager image, MIG-capable hardware); vGPU, KubeVirt passthrough, confidential computing, and GPUDirect Storage are feature-gated and not in the default package

Validation tiers

Not every configuration the operator accepts has been validated end-to-end. This table records the validation level for this release.

ConfigurationHardware / OSValidation level
Precompiled driver, default flow, reboot survivalSUSE Linux Micro 5.5 (kernel 5.14.21-150500.55.166), RTX 4090 + Tesla T4e2e validated
Preinstalled host driver (driver.enabled=false)RHEL 9.6, Tesla P100, host driver 570.124.06e2e validated
OLM skipRange upgrade to v26.3.3-1as above, with ClusterPolicy presentvalidated (smooth upgrade, driver not rebuilt)
Cross-driver-version migration (570 ↔ 580)as abovevalidated (GPU has a loss window during module reload)
Upstream driver (mutable OS, nvcr.io)Ubuntu / RHELnot validated — configurable, declared out of scope for this release
MIG partitioningMIG-capable GPUnot validated this release — separate follow-up task

An accepted configuration that is not listed as validated is supported at the parameter level only; confirm it against your environment before production use.

Precompiled driver image compatibility

A precompiled driver image is kernel-specific: it loads only on the exact kernel it was built for. The tuple tag encodes the driver version (including the Alauda revision), the complete kernel release, and the OS:

<driver-version>-<kernel-release>-<os>
# example: 580.126.09.1-5.14.21-150500.55.166-default-sle-micro5.5

The product package ships the tuple validated in this release (R580, SUSE Linux Micro 5.5, kernel 5.14.21-150500.55.166). Additional tuples are published on Docker Hub; import the one that matches your node as described in Configure the driver mode.

The driver version (for example R580 or R570) is a choice you make for your workloads. The kernel and OS are node facts and must match exactly — a tuple built for a different kernel does not make that kernel compatible. If your node kernel has no published tuple, contact Alauda Customer Support with uname -r and the GPU model for an on-demand build; tuples are not published speculatively for untested kernels.

Known limits

  • In-place migration from the previous standalone cluster-plugin delivery model to the operator is not supported. Uninstall the standalone plugins before installing the operator.
  • A driver version change reloads kernel modules on the node; GPU workloads on that node lose GPU access during the reload. Plan it as a node maintenance window.
  • The upstream mutable-OS driver mode builds kernel modules on the node; air-gapped clusters need an internal kernel-headers source (driver.repoConfig).
  • MIG is not validated on MIG hardware in this release and requires a user-supplied mig-manager image.
  • Vendor hardware health, firmware, and low-level driver tuning are outside this product document and should follow NVIDIA guidance.