GPU instances autoscaling

The CAST AI autoscaler supports running your workloads on GPU-optimized instances.
This guide will help you configure and run it in 5 minutes.

Supported providers

ProviderGPUs supported
AWS EKSNVIDIA (AMD coming soon)
GCP GKEcoming soon
EKS KOPScoming soon
Azure AKScoming soon


GPU Driver

A GPU-specific driver should be installed on the cluster to run GPU workloads. CAST AI verifies and ensures that the required driver is available on the cluster before provisioning the required nodes.

How to install the NVIDIA driver

  1. Onboard a cluster to CAST AI by providing additional variable INSTALL_NVIDIA_DEVICE_PLUGIN=true

  2. Install it from NVIDIA repository:

    helm repo add nvdp
    helm repo update
    noglob helm upgrade -i nvdp nvdp/nvidia-device-plugin -n castai-agent \
        --set-string nodeSelector."nvidia\.com/gpu"=true \
        --set \
  3. Use your own plugin. CAST AI does a plugin compatibility check with a new node before provisioning it, so CAST AI should detect a plugin in order to perform the check. Plugin will be detected by CAST AI if one of these conditions are honored:

    • plugin daemon set name pattern is *nvidia-device-plugin*
    • plugin daemon set has label nvidia-device-plugin: "true"

Workload configuration

To request a node that has an attached GPU, workload should:

  • define (at least) GPU limits in the workload resources:

        cpu: 1
        memory: 1Gi 1
        memory: 1Gi 1
  • add a toleration for GPU node (toleration is required because CAST AI adds a taint on GPU nodes so that these nodes could be used only by workloads that truly require GPUs):

        - key: ""
          operator: Exists