During scale up, the cluster autoscaler is triggering unnecessary scale ups. This issue was only noticed when using GPUs. An issue has been filed in the cluster autoscaler project https://github.com/kubernetes/autoscaler/issues/3164.