AMD GPU Operator | Hardware accelerators

About the AMD GPU Operator
Installing the AMD GPU Operator
Testing the AMD GPU Operator

AMD Instinct GPU accelerators combined with the AMD GPU Operator within your OKD cluster lets you seamlessly harness computing capabilities for machine learning, Generative AI, and GPU-accelerated applications.

This documentation provides the information you need to enable, configure, and test the AMD GPU Operator. For more information, see AMD Instinct™ Accelerators.

About the AMD GPU Operator

The hardware acceleration capabilities of the AMD GPU Operator provide enhanced performance and cost efficiency for data scientists and developers using Red Hat OpenShift AI for creating artificial intelligence and machine learning (AI/ML) applications. Accelerating specific areas of GPU functions can minimize CPU processing and memory usage, improving overall application speed, memory consumption, and bandwidth restrictions.

Installing the AMD GPU Operator

As a cluster administrator, you can install the AMD GPU Operator by using the OpenShift CLI and the web console. This is a multi-step procedure that requires the installation of the Node Feature Discovery Operator, the Kernel Module Management Operator, and then the AMD GPU Operator. Use the following steps in succession to install the AMD community release of the Operator.

Next steps

Install the Node Feature Discovery Operator.
Install the Kernel Module Management Operator.
Install and configure the AMD GPU Operator.

Testing the AMD GPU Operator

Use the following procedure to test the ROCmInfo installation and view the logs for the AMD MI210 GPU.

Procedure

Create a YAML file that tests ROCmInfo:

$ cat << EOF > rocminfo.yaml

apiVersion: v1
kind: Pod
metadata:
 name: rocminfo
spec:
 containers:
 - image: docker.io/rocm/pytorch:latest
   name: rocminfo
   command: ["/bin/sh","-c"]
   args: ["rocminfo"]
   resources:
    limits:
      amd.com/gpu: 1
    requests:
      amd.com/gpu: 1
 restartPolicy: Never
EOF

Create the rocminfo pod:

$ oc create -f rocminfo.yaml

Example output

apiVersion: v1
pod/rocminfo created

Check the rocmnfo log with one MI210 GPU:

$ oc logs rocminfo | grep -A5 "Agent"

Example output

HSA Agents
==========
*******
Agent 1
*******
  Name:                    Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz
  Uuid:                    CPU-XX
  Marketing Name:          Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz
  Vendor Name:             CPU
--
Agent 2
*******
  Name:                    Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz
  Uuid:                    CPU-XX
  Marketing Name:          Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz
  Vendor Name:             CPU
--
Agent 3
*******
  Name:                    gfx90a
  Uuid:                    GPU-024b776f768a638b
  Marketing Name:          AMD Instinct MI210
  Vendor Name:             AMD

Delete the pod:

$ oc delete -f rocminfo.yaml

Example output

pod "rocminfo" deleted