Run workflows on specified GPU and AMD ECS instances - Container Service for Kubernetes

Workflow clusters let you pin each workflow step to a specific Elastic Compute Service (ECS) instance type — GPU-accelerated or AMD-based — by adding a single annotation to the pod template. Use this when your workloads require specific hardware, such as GPU computation, video encoding, or high-memory processing.

Prerequisites

Before you begin, ensure that you have:

A workflow cluster in Container Service for Kubernetes (ACK)
Permissions to create and submit Argo Workflows

How it works

Add the k8s.aliyun.com/eci-use-specs annotation under metadata.annotations in the template that requires custom hardware. The annotation value is the ECS instance type, such as ecs.gn5i-c4g1.xlarge.

All elastic container instances created for that template step are scheduled onto the specified ECS instance type.

Supported instance types

GPU-accelerated instance types

GPU-accelerated elastic container instances support NVIDIA driver version 460.73.01 and CUDA Toolkit 11.2. For available CUDA images, see NVIDIA CUDA on Docker Hub.

Instance family	GPU	Example instance type
gn6v	NVIDIA V100	ecs.gn6v-c8g1.2xlarge
gn6i	NVIDIA T4	ecs.gn6i-c4g1.xlarge
gn5	NVIDIA P100	ecs.gn5-c4g1.xlarge
gn5i	NVIDIA P4	ecs.gn5i-c2g1.large

For the full list, see Instance families.

AMD-based instance types

AMD-based elastic container instances run on AMD EPYC™ ROME processors and use the SHENLONG architecture to minimize virtualization overhead. They work well for video encoding and decoding, large packet throughput, web frontend servers, MMO (massively multiplayer online) game frontends, and DevOps application development and testing.

Instance family	Type	Example instance type
g7a, g6a	General-purpose	ecs.g7a.large, ecs.g6a.large
c7a, c6a	Compute-optimized	ecs.c7a.large, ecs.c6a.large
r7a, r6a	Memory-optimized	ecs.r7a.large, ecs.r6a.large

For the full list, see Instance families.

Run a workflow on a GPU-accelerated instance

The following example schedules the whalesay template on a gn5i GPU instance by setting k8s.aliyun.com/eci-use-specs in the template's annotations.

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: hello-world-
spec:
  entrypoint: whalesay
  templates:
    - name: whalesay
      metadata:
        annotations:
          k8s.aliyun.com/eci-use-specs: ecs.gn5i-c4g1.xlarge  # GPU-accelerated instance type
      container:
        image: docker/whalesay
        command: [cowsay]
        args: ["hello world"]

Run a workflow on an AMD-based instance

The following example schedules the whalesay template on a c6a AMD instance.

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: hello-world-
spec:
  entrypoint: whalesay
  templates:
    - name: whalesay
      metadata:
        annotations:
          k8s.aliyun.com/eci-use-specs: "ecs.c6a.xlarge"  # AMD-based instance type
      container:
        image: docker/whalesay
        command: [cowsay]
        args: ["hello world"]

What's next

For the full list of ECS instance families and their specifications, see Instance families.