NVIDIA AI Enterprise 9.0 Release Notes#

These release notes describe NVIDIA® AI Enterprise Infrastructure Release 9.0. They list the supported infrastructure software components and versions, compatibility and support information, and links to per-component release notes.

Note

NVIDIA AI Enterprise Infra 9.0 is a Production Branch (PB) following the standard release cadence with feature additions and quarterly support. For branch lifecycle, support windows, and Long-Term Support Branch options, refer to the NVIDIA AI Enterprise Lifecycle Policy.

Key Features and Enhancements#

  • NVIDIA Vera Rubin NVL72: NVIDIA Vera Rubin NVL72 is a newly supported platform for bare metal. Refer to the Support Matrix.

  • NVIDIA Data Center GPU Driver 620.70.06: New R620 production driver branch (from 595.112.07 in 8.3). Refer to the 620.70.06 release notes for fixes and platform-support details.

  • NVIDIA Fabric Manager 2.0.0.rel102: Fabric Manager now ships as a separate package for the NVIDIA Virtual GPU Manager and is versioned on its own release line rather than the driver branch. It remains included with the NVIDIA Data Center GPU Driver. Install the Virtual GPU Manager first. Refer to NVIDIA Fabric Manager on HGX Servers.

  • NVIDIA vGPU™ for Compute updates: The NVIDIA Virtual GPU Manager (620.70.04, from 595.112.08 in 8.3) and the NVIDIA vGPU for Compute Guest Driver (Linux 620.70.06, from 595.112.07 in 8.3; Windows 621.14, from 597.30 in 8.3) refresh the full vGPU stack in a coordinated release, including performance monitoring, live migration, and scheduling optimizations:

    • Support for vGPU live migration when NVIDIA CUDA® Toolkit debuggers or profilers are enabled

    • Support for the Microsoft Compute Driver Model (MCDM)

    • Support for notifying guest vGPU for Compute VMs when a GPU reset is required

    • Support for GPU performance monitoring (GPM) metrics through NVAPI and NvWMI (NVIDIA Windows Management Instrumentation)

  • Azure Local support for NVIDIA vGPU for Compute: NVIDIA vGPU for Compute now adds support for the Microsoft Azure Local hypervisor on NVIDIA RTX PRO 6000 Blackwell Server Edition. Supported versions are Microsoft Azure Local 12.2609.1003.7 and Microsoft Azure Local 12.2510.

  • NVIDIA NIM™ Operator 3.1.4: Updated from 3.1.2 in 8.3. Refer to the NIM Operator 3.1.4 release notes for fixes and enhancements.

What Is Included in NVIDIA AI Enterprise Infra 9.0#

Complete list of infrastructure components with versions and documentation links:

Table 15 Supported Infrastructure Software#

Component

Description

Version

NVIDIA Data Center GPU Driver

Provides hardware support for NVIDIA GPUs.

620.70.06

NVIDIA DOCA™ Driver for Networking

Provides hardware support for NVIDIA BlueField® DPUs and SuperNICs. Installing DOCA on the host provides all necessary drivers and tools to manage BlueField and ConnectX® devices.

3.5.0

NVIDIA Fabric Manager

Manages NVSwitch™ fabric to enable high-bandwidth, low-latency GPU-to-GPU communication for multi-GPU AI workloads. Starting with NVIDIA AI Enterprise Infra 9.0, Fabric Manager is a separate package for the NVIDIA Virtual GPU Manager and is versioned on its own release line rather than the driver branch. It remains included with the NVIDIA Data Center GPU Driver.

2.0.0.rel102

NVIDIA DOCA Microservices

Infrastructure acceleration and offload services for NVIDIA BlueField, enabling accelerated networking, storage, and security workloads.

3.5.0

NVIDIA Virtual GPU Manager

GPU driver deployed alongside the hypervisor in virtualized environments. Enables multi-tenant GPU sharing, live migration, and monitoring.

620.70.04

NVIDIA vGPU for Compute Guest Driver [1]

GPU driver deployed in the guest VM to enable multiple VMs to have simultaneous, direct access to a single physical GPU.

NVIDIA Container Toolkit

Enables GPU-accelerated containers by providing runtime components and utilities for container engines (Docker, containerd, CRI-O)

1.20.1

NVIDIA Run:ai

Provides a Kubernetes-native orchestration and management platform that maximizes GPU utilization for AI workloads through advanced scheduling and resource management.

3.0

NVIDIA DPU Operator (DPF)

Enables cluster administrators to automate provisioning, orchestration, and lifecycle management of BlueField DPUs and DOCA Microservices to enable DPU-accelerated North-South networking in Kubernetes.

26.8.0

NVIDIA GPU Operator

Simplifies deployment of NVIDIA AI Enterprise by automating the management of all NVIDIA software components needed to provision GPUs in Kubernetes, including the NVIDIA DRA driver. The driver is a supported component when deployed through the GPU Operator.

26.7.2

NVIDIA Network Operator

Simplifies deployment of high-speed networking by automating management of NVIDIA ConnectX NICs and SuperNICs required to optimize East-West traffic and RDMA transfers in Kubernetes.

26.7.1

NVIDIA NIM™ Operator

Enables cluster administrators to operate the software components and services required to run LLM, embedding, and other models using NVIDIA NIM microservices in Kubernetes.

3.1.4

NVIDIA Base Command Manager (BCM)

Cluster management and provisioning tool for NVIDIA DGX™ systems.

11.34.0

Compatibility and Support#

Support Matrix

Use the Interactive Support Matrix to compare NVIDIA® AI Enterprise infrastructure compatibility across releases. The web tool lets you:

  • Filter by deployment type, operating system, hypervisor, or orchestration platform.

  • Search by GPU architecture, platform, Kubernetes distribution, or cloud provider.

  • View release badges and footnote tooltips for each configuration.

For deep linking, printing, or offline reference, the same information is also available as the static Support Matrix. Both forms list supported GPUs, operating systems, hypervisors, orchestration platforms, cloud instance types, and networking hardware.

Key support boundaries

The following limits are the most common deployment blockers. For the complete matrix and all footnotes, refer to the Support Matrix.

  • NVIDIA vGPU™ for Compute on Linux with KVM is supported for single-node deployments only.

  • NVIDIA Grace Hopper and NVIDIA Grace Blackwell platforms support bare metal, GPU Passthrough, and NVIDIA vGPU for Compute. NVIDIA vGPU for Compute requires Linux with KVM on an ARM64 host.

  • GPU Operator with KubeVirt or OpenShift Virtualization on NVIDIA HGX™ Hopper and Blackwell platforms requires GPU Passthrough.

Lifecycle and Compatibility Explorer

Use the Interactive Lifecycle and Compatibility Explorer on the NVIDIA® AI Enterprise Lifecycle Policy documentation to:

  • Query by Infrastructure Branch, by release, or by component type and version.

  • Run a full-stack check from a GPU driver version to validate compatibility and plan upgrades.

Footnotes