Red Hat has launched OpenShift 4.21 with Dynamic Resource Allocation for GPUs, which allows high-end GPUs to be prioritized for AI training. These resources can also be scaled down completely to save money. The release also adds autoscaling to zero for hosted control planes and cross-cluster VM migration without downtime.
OpenShift 4.21 addresses a fundamental problem that AI teams face daily: GPU allocation that does not match their actual needs. Traditionally, teams simply requested a GPU and hoped it would meet their requirements. With the new Dynamic Resource Allocation, users specify exactly what they need, for example, “a GPU with at least 40GB VRAM.” The scheduler queries hardware attributes directly via common expression language to find the right resources.
This eliminates manual node labeling. The system reads hardware capabilities and automatically matches them to workload requirements. This feature does require a vendor-provided operator or driver with DRA support.
Cost optimization through autoscaling
Hosted control planes get native VerticalPodAutoscaler integration. Control plane components scale automatically based on real-time memory consumption rather than static estimates. In addition, control planes can now scale to zero during inactivity. The configuration and state are preserved, and they resume automatically when needed.
NodePools follow the same pattern and scale to zero nodes in development and test environments. This makes Hosted Control Planes the most cost-efficient way to run OpenShift. The feature ensures that the hosted control plane remains operational in standby mode.
Virtualization is maturing
Cross-cluster live migration allows administrators to move running VMs between different OpenShift clusters without downtime. Administrators can now perform cluster maintenance, rebalance resources across regions, or migrate workloads to newer hardware without service interruption.
IPv6-only control plane and secondary network support is now generally available. This is an important step for organizations running out of IPv4 addresses. It allows you to roll out OpenShift clusters and virtualized workloads in modern, IPv6-native environments. Complex Network Address Translation workarounds are no longer necessary.
OpenShift Virtualization on Google Cloud bare metal enables organizations to run VMs directly on dedicated hardware. This deployment model is critical for performance-sensitive workloads that require direct access to physical CPU features and hardware acceleration, such as low-latency databases.
Improved UI for network management
The enhanced virtualization UI guides administrators to the correct network configurations while maintaining advanced control. Administrators can now create secondary ClusterUserDefinedNetworks with localnet topology. Built-in guardrails prevent accidental deletion of UDN-derived NetworkAttachmentDefinitions.
Red Hat OpenShift Lightspeed, the virtual AI assistant, is integrated with the OpenShift Virtualization user interface. Virtualization administrators no longer need to switch between interfaces or manually upload files. They now get in-context, AI-powered insights for troubleshooting virtual machine errors.
OpenShift 4.20 previously introduced post-quantum encryption and LeaderWorkerSet for AI workloads. Version 4.21 builds on this with practical improvements for everyday use.