GPU Worker Nodes Early Availability

DigitalOcean Kubernetes (DOKS) is a managed Kubernetes service. Deploy Kubernetes clusters with a fully managed control plane, high availability, autoscaling, and native integration with DigitalOcean Load Balancers and volumes. DOKS clusters are compatible with standard Kubernetes toolchains and the DigitalOcean API and CLI.

GPU worker nodes are now in early availability for select DOKS users. You can either create a new cluster or add a GPU node pool on an existing cluster on versions 1.30.4-do.0, 1.29.8-do.0, 1.28.13-do.0, and later.

GPU worker nodes are built on GPU Droplets, which are powered by NVIDIA’s H100 GPUs.

Using GPU worker nodes in your cluster, you can:

• Experiment and develop AI/ML applications in containerized environments
• Run distributed AI workloads on Kubernetes
• Scale AI inference services

You do not need to specify a Runtime Class to run GPU workloads, which makes it easier to set up GPU workloads.

GPU Droplet Details

GPU Droplets have NVIDIA H100 GPUs in either a single GPU or 8 GPU configuration, and come with two different kinds of storage:

• Boot disk: A local, persistent disk on the Droplet to store data for software like the operating system and ML frameworks.

• Scratch disk: A local, non-persistent disk to store data for staging purposes, like inference and training.

The following table summarizes additional specifications:

Pricing

GPU worker nodes are priced per second at the same price as the GPU Droplets. See DOKS node pools pricing to learn more about pricing.

For reservation and contract pricing, contact your sales representative or Customer Success Manager, or send a request using the H100 GPU Worker Nodes form.

Availability

GPU worker nodes for DOKS are currently available in DigitalOcean’s TOR1 datacenter. We plan to support additional datacenters in the near future.

Limits

• You cannot currently scale the GPU node pools down to zero.

• You need to monitor GPU usage and manage scaling manually.

• We do not currently support creating or adding GPU worker nodes using the DigitalOcean Control Panel.

Getting Started

Add a GPU Worker Node Using Automation

You can use the DigitalOcean CLI or API to create a new cluster or add a GPU node pool on an existing cluster.

Note

In rare cases, it can take several hours for a GPU Droplet to provision. If you have an unusually long creation time, open a support ticket.

doctl kubernetes cluster create gpu--cluster --region tor1 --version 1.30.4-do.0 --node-pool "name=gpu-worker-pool;size=gpu-h100x1-80gb;count=3"

doctl kubernetes cluster node-pool create gpu-cluster --name gpu-worker-pool-1 --size gpu-h100x1-80gb --count 4

curl --location 'https://api.digitalocean.com/v2/kubernetes/clusters' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer $DIGITALOCEAN_TOKEN' \
--data '{
    "name": "gpu--cluster",
    "region": "tor1",
    "version": "1.30.4-do.0",
    "node_pools": [
        {
            "size": "gpu-h100x1-80gb",
            "count": 3,
            "name": "gpu-worker-pool"
        }
    ]
}'

curl --location --request POST 'https://api.digitalocean.com/v2/kubernetes/clusters/{cluster_id}/node_pools' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer $DIGITALOCEAN_TOKEN' \
--data '{
    "node_pools": [
        {
            "size": "gpu-h100x1-80gb",
            "count": 4,
            "name": "new-gpu-worker-pool"
        }
}',

DigitalOcean applies the following labels and taint to the GPU worker nodes:

You can use the labels and taint in your workload specification to schedule pods that match as shown next.

Run a GPU Workload

Once you have a GPU worker node on your cluster, you can run GPU workloads. You can use a configuration spec, similar to the sample pod spec, for your actual workloads. The following spec shows how to create a pod that runs NVIDIA’s CUDA image and uses the labels and taint for GPU worker nodes:

    
        
            
apiVersion: v1
kind: Pod
metadata:
  name: gpu-workload
spec:
  restartPolicy: Never
  nodeSelector:
    doks.digitalocean.com/gpu-brand: nvidia
  containers:
    - name: cuda-container
      image: nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda10.2
  tolerations:
    - key: nvidia.com/gpu
      operator: Exists
      effect: NoSchedule

        
    

Optional Additional Configuration

• DigitalOcean installs and manages the drivers required to enable GPU support on the GPU worker nodes. However, for GPU discovery, health checks, configuration of GPU-enabled containers, and time slicing, you need an additional component called NVIDIA device plugin for Kubernetes. You can configure the plugin and deploy it using helm, as described in the README file of the GitHub repository.

• For monitoring your cluster using Prometheus, you need to install NVIDIA DCGM Exporter.

Feedback

For support or troubleshooting, open a support ticket.

For feedback or questions about the GPU worker nodes offering, contact your account representative.