GKE Autopilot Clusters Now Support Managed DRANET with GPUs and TPUs
Google Kubernetes Engine (GKE) Autopilot clusters now support managed DRANET, enabling the allocation of networking resources for Pods, including those with TPUs and Remote Direct Memory Access (RDMA) for GPUs. This feature simplifies the setup for users by automating node management and security configurations within Autopilot. The enhancement is particularly relevant for engineers and architects working with accelerators for machine learning and high-performance computing workloads. The documentation outlines a step-by-step guide for deploying Autopilot clusters and configuring ComputeClass and ResourceClaimTemplate resources to leverage this new networking capability.
- →GKE Autopilot supports managed DRANET for GPUs and TPUs
- →Configuration guidance for Autopilot with managed DRANET
- →Prerequisites for GKE Autopilot managed DRANET setup
Features (2) ›
- GKE Autopilot supports managed DRANET for GPUs and TPUs
Google Kubernetes Engine (GKE) Autopilot clusters now support managed DRANET, which allows for the request and allocation of networking resources for Pods, including those that require TPUs and RDMA for GPUs. This feature aims to simplify the configuration of advanced networking for accelerator-equipped workloads within the managed Autopilot environment.
- Configuration guidance for Autopilot with managed DRANET
The release includes detailed instructions for setting up GKE Autopilot clusters with managed DRANET, covering the creation of VPC networks, custom ComputeClasses for specific accelerator types (TPU v6e or NVIDIA B200 GPUs), and ResourceClaimTemplates for RDMA or non-RDMA devices. Workload deployment examples are provided, showcasing how to reference these resources to ensure correct networking setup for AI and ML inference.
Notes (1) ›
- Prerequisites for GKE Autopilot managed DRANET setup
To deploy GKE Autopilot clusters with managed DRANET, users need to create a Virtual Private Cloud (VPC) and configure specific environment variables. These include project ID, region, cluster name, network and subnetwork details, reservation URL for dedicated resources, and a Hugging Face access token for model downloads.
https://cloud.google.com/blog/topics/developers-practitioners/autopilot-clusters-with-gke-managed-dranet-gpus-and-tpus/
Related releases
- Google Cloud C4N instances achieve GA for optimized network and storage I/O Google Cloud Blog ·
- Google Cloud Recognized as Leader in Gartner AI Infrastructure Magic Quadrant Google Cloud Blog ·
- Report: Agentic AI requires significant infrastructure upgrades Google Cloud Blog ·
- Google Cloud: New AI, Serverless, and Infrastructure Updates Google Cloud Blog ·
- Google Cloud enables AI-defined vehicles with Android and Bigtable Google Cloud Blog ·
- k8s-aibom open-sourced for automated AI Bill of Materials on GKE Google Cloud Blog ·