Alauda Ascend NPU
  • English

    • Compatibility

    Use this page to check whether an NPU Operator deployment matches the supported product scope.

    TOC

    Product support scopePrecompiled driver image compatibilityKnown limits

    Product support scope

    ItemSupported scope
    ACPv4.0-v4.3
    HardwareAscend 910B, Ascend 310P
    ArchitectureARM
    Driver lifecyclePre-installed driver, operator-managed ordinary OS install, or operator-managed immutable OS / Alauda OS precompiled install
    Runtime integrationAscend Device Plugin allocation with the ascend RuntimeClass backed by ascend-docker-runtime; the admission webhook injects the RuntimeClass by default
    MonitoringNPU Exporter; ServiceMonitor when the monitoring namespace exists
    Node discoveryAlauda Build of Node Feature Discovery plus NPU feature discovery components delivered by the target package
    Optional Ascend componentsMindIO TFT, MindIO ACP, NodeD, ClusterD, and Resilience Controller only when enabled and validated for the target release

    Precompiled driver image compatibility

    Precompiled Driver image tags are kernel-specific. NPU Operator derives a tag from the HDK version, detected chip, and complete kernel release after removing the architecture suffix:

    <HDK>-<chip>-<kernel-release-without-architecture>

    For the published openEuler tag layout, .oe in the kernel release is normalized to -oe. Other kernel releases remain unchanged. The resulting image must still match the node operating system; the tag algorithm does not make a different OS or kernel compatible.

    If your node kernel has no matching driver image tag, contact Alauda Customer Support. Do not treat a different kernel tag as compatible.

    See the Alauda Ascend Driver tags on Docker Hub for the published image inventory. The installation guide provides concrete CTyunOS 310P and openEuler 910B examples. Publication of another tag does not by itself assert product support or validation for your cluster.

    Known limits

    • In-place upgrade from the old v1.1.x cluster plugin delivery model to the OperatorHub / OLM delivery model is not supported. Uninstall the old cluster plugin before installing the operator.
    • Driver upgrade and chip recovery can require node reboot.
    • Automatic HCCN device IP initialization is not supported in v26.6.0. Configure 910-family HCCN addresses before workload verification; the validated HCCN Bootstrap profile is observe-only and does not rewrite existing addresses.
    • The Ascend runtime handler configuration can be lost if a host containerd package upgrade replaces /etc/containerd/config.toml; restart the runtime integration DaemonSet to re-apply it.
    • Vendor hardware health, BIOS, firmware, and low-level driver tuning are outside this product document and should follow vendor guidance.