Upgrading from 8.x to 9.0#
Use this checklist when upgrading existing 8.x deployments to 9.0. Run each step in a maintenance window and validate before proceeding.
Note
NVIDIA® Data Center GPU Manager (DCGM) is not supported in vGPU™ environments (hypervisor hosts running Virtual GPU Manager, and vGPU guest VMs). Do not use DCGM metrics as a post-upgrade health check on those paths. Refer to Limitations.
Component Version Delta#
Compare infrastructure versions across 8.2, 8.3, and 9.0 before you upgrade. Find your current release in the column headings and read across to the 9.0 version. Components that keep the same version in all three releases are omitted. For full component descriptions, refer to the Release Notes. For platform and OS support, refer to the Support Matrix.
Component |
8.2 |
8.3 |
9.0 |
|---|---|---|---|
NVIDIA Data Center GPU Driver |
595.91.07 |
595.112.07 |
620.70.06 |
NVIDIA DOCA™ Driver for Networking |
3.4.0 |
3.5.0 |
3.5.0 |
NVIDIA Fabric Manager |
595.91.07 |
595.112.07 |
2.0.0.rel102 |
NVIDIA DOCA Microservices |
3.4.0 |
3.5.0 |
3.5.0 |
NVIDIA Virtual GPU Manager |
595.91.04 |
595.112.08 |
620.70.04 |
NVIDIA vGPU™ for Compute Guest Driver (Linux) |
595.91.07 |
595.112.07 |
620.70.06 |
NVIDIA vGPU™ for Compute Guest Driver (Windows) |
596.86 |
597.30 |
621.14 |
NVIDIA Container Toolkit |
1.19.1 |
1.20.1 |
1.20.1 |
NVIDIA Run:ai |
2.26 |
3.0 |
3.0 |
NVIDIA DPU Operator (DPF) |
26.4.0 |
26.8.0 |
26.8.0 |
NVIDIA GPU Operator |
26.3.3 |
26.7.2 |
26.7.2 |
NVIDIA Network Operator |
26.4.1 |
26.7.1 |
26.7.1 |
NVIDIA NIM™ Operator |
3.1.2 |
3.1.2 |
3.1.4 |
NVIDIA Base Command Manager (BCM) |
11.33.1 |
11.34.0 |
11.34.0 |
Upgrade Sequence#
Complete the following steps in order during a maintenance window. Validate each step before you continue to the next.
Step |
Action |
Reference |
|---|---|---|
|
Read the component version delta in the table above and review per-component release notes for breaking changes, deprecations, and feature additions. |
|
|
Confirm hardware, hypervisor (for virtualized deployments), and operating-system compatibility against the support matrix. |
|
|
In virtualized environments, snapshot or back up VMs and capture current driver, operator, and licensing configuration before upgrading. Document the rollback path for each component. |
N/A |
|
Depending on your deployment type, that is, bare metal or virtualized, upgrade the NVIDIA Data Center GPU Driver (applies to bare metal deployments), Virtual GPU Manager (only applies to vGPU for Compute deployments), and DOCA™ Driver. Fabric Manager remains included with the NVIDIA Data Center GPU Driver. For the Virtual GPU Manager, upgrade Fabric Manager as a separate package after the Virtual GPU Manager, and on VMware vSphere ESXi reboot the host after you install it. |
|
|
Upgrade NVIDIA GPU Operator, Network Operator, NIM™ Operator, DPU Operator (DPF), and NVIDIA Run:ai (self-hosted) to the versions in the component version delta table above. NVIDIA Run:ai self-hosted moves from 2.26 to 3.0 when you upgrade from 8.2 (8.3 already ships 3.0). This is a major version change with its own prerequisites and supported upgrade paths, so plan it as a separate task within this step and follow the NVIDIA Run:ai upgrade guide. NVIDIA Run:ai SaaS upgrades are managed by NVIDIA and require no customer-side upgrade. |
Component version delta above, NVIDIA Run:ai Upgrade Guide |
|
For virtualized deployments, upgrade NVIDIA Datacenter driver (GPU passthrough), NVIDIA vGPU for Compute Guest Driver (vGPU deployments), or NVIDIA Container Toolkit in tenant VMs. |
|
|
Confirm each licensed vGPU VM reaches the NVIDIA License System and shows |
|
|
Run a representative CUDA® or AI/ML workload to confirm performance, feature parity, and operator-managed scheduling behavior. |
N/A |
For lifecycle policy, branch support windows, and migration windows beyond 8.x, refer to the NVIDIA AI Enterprise Lifecycle Policy.