This blog post details setting up Google Kubernetes Engine (GKE) Autopilot clusters with managed DRANET to support GPUs and TPUs. GKE Autopilot simplifies Kubernetes management by handling nodes, scaling, and security, while managed DRANET enables requesting network resources for Pods, including those for TPUs and RDMA-enabled GPUs. The setup involves creating a Virtual Private Cloud (VPC), deploying an Autopilot cluster, and then configuring custom ComputeClasses and ResourceClaimTemplates.Configure essential variables such as region, cluster name, network, subnetwork, reservation URL, and Hugging Face token before proceeding. First, an Autopilot cluster is deployed, providing a fully managed environment. Next, a custom ComputeClass is created to specify the accelerator type (GPU or TPU), optionally referencing a reservation for resource allocation.Following this, a ResourceClaimTemplate is established, differentiating between RDMA-enabled GPUs (using deviceClassName: mrdma.google.com) and non-RDMA TPUs (using netdev.google.com). Finally, a workload is deployed, referencing both the custom ComputeClass and the ResourceClaimTemplate. This crucial step triggers GKE Autopilot to provision the specified node type and configure managed DRANET networking.The resource claim then acts as a bridge, directly binding the Pods to the accelerators on the provisioned nodes, ensuring the correct networking setup for the workload. This comprehensive process applies uniformly for both GPU and TPU deployments. Additional resources are provided for further exploration.