Introducing support for GPU workloads and even larger Pods in GKE Autopilot