Docs

Volcengine Kubernetes Engine

GPU-accelerated computing instance

Create an instance by specifying an instance family (GPU resource).

VCI allows you to specify instance families by using specific annotations, and create GPU-type VCIs by specifying resources.limits.

Usage limits

The supported regions and availability zones vary based on different GPU-accelerated instance families. For details, see Regions and availability zones.

When you create a GPU Accelerated instance, different configurations match different VCI instance specifications. For the detailed specification matching logic, see Instance specification normalization.

Prerequisites

A cluster for VCI applications has been created. For details, see Using VCI in VPC-CNI clusters and Using VCI in Flannel clusters.

A VPC and its subnet have been created. For details, see Creating VPCs.

Creating instances by using Kubernetes

You can create GPU-type VCI pods by adding annotations and specifying resources.limits during the creation of pods or deployments, including Deployments, StatefulSets, Jobs, and CronJobs.

The following table describes the annotation.

Annotation key	Annotation sample value	Description	Required
vci.vke.volcengine.com/preferred-instance-family	vci.ini2	The instance families of the GPU-accelerated type. You also need to specify the resources.limits of the pod. VCI will match instance specifications based on the specification matching logic. tip For the value description of the annotation used to create a general-purpose instance, see Creating general-purpose instances by specifying container vCPUs and memory. Billing is performed based on the matched instance specification. For billing descriptions, see Product billing. You can specify multiple GPU-accelerated instance families (annotation values). Separate multiple instance families with commas (,), and determine the priority by the order in which they are entered. If the higher-priority instance family has no available resources, the lower-priority instance family will be used to create instances.	Required for this scenario
vci.vke.volcengine.com/gpu-driver-version	tesla-470.129.06	Specify the driver version of the GPU-accelerated instance. The annotation value is in the following format: tesla-{NVIDIA driver version}. The available {NVIDIA driver version} varies depending on the GPU-accelerated specifications specified in preferred-instance-family. For details, see Supported GPU driver versions. tip If you do not specify this annotation, the system automatically specifies the default NVIDIA driver version that corresponds to the selected GPU-accelerated instance family.	Optional

The following sample code shows how to create a pod.

tip

The GPU settings under Pod.spec.container[].resources must align with native Kubernetes capabilities and limitations.

Use <vendor>.com/gpu to specify GPU resources.

The limits option is required.

Both limits and requests can be configured, but their values must be identical.

If the requests option is left empty, Kubernetes will use the limits value as the requests value by default.

GPU cannot be shared and overcommitted.

The number of GPUs must be an integer.

Using the console

Install the nvidia-device-plugin add-on in the VCI cluster. For details, see Installing add-ons.

Specify the following parameters when you create a workload.

Basic information: Select Volcengine Container Instance for the Computing power type parameter.

Container configuration: Specify the GPU, memory, and CPU resources in the ResourceQuota section.

Specify the parameter in the red box in the following figure. Set the GPU resource type parameter to Nvidia. Specify other parameters based on your business requirements.

tip

The mGPU resource type is not supported in VCI scenarios.

Advanced configurations: Configure the instance specifications in the VCI specification settings section. You can use the default instance family or specify a custom instance family.

When specifying instance families, you can select multiple families under the same instance type. Their priority is determined by the order in which they are selected. For details, see Specific specifications. If resources are unavailable in a higher-priority family, VCI uses a lower-priority family to create the instance.

Last updated: 2026.09.03 15:56:09