Releases · NVIDIA/gpu-operator

New Features

Support for NVIDIA Data Center GPU Driver version 470.82.01.
MIG Manager support with Preinstalled GPU Drivers on the node.
Entitlement free driver builds using DriverToolkit images on RedHat OpenShift 4.9.
Support for enabling GPUDirect RDMA with Preinstalled MOFED drivers on the node using driver.rdma.useHostMofed option.
Upgrade support of GPU Operator and Operand Pods using Operator Lifecycle Manager(OLM) on RedHat OpenShift.

Automatic detection of default runtime used in the cluster and deprecate operator.defaultRuntime parameter.
Installation of GPU Operator and its Operands into a single user specified Namespace.
Automatic detection and unload of nouveau driver prior to installing NVIDIA GPU Driver on each node.
Option to mount a ConfigMap of self-signed certificates into driver container using driver.certConfig.name parameter. This can be
used for SSL connections to Proxy server or private package repositories.

DCGM Exporter in CrashLoopBackOff as it cannot connect to DCGM port on the same node.

Upgrades from v1.8.x to v1.9.x are not supported due to change in default Namespace gpu-operator-resources used earlier.
Starting from v1.9.x user can specify custom Namespace for both GPU Operator and Operand Pods to be installed. User would need
to un-install old version prior to installing v1.9.x version of GPU Operator.
Collection of GPU metrics in MIG mode is not supported with 470+ drivers.