Try NVIDIA AICR for GPUs · ConfigHub Workshop

Compare a GPU workload with supplied Node facts with direct cub commands or an assistant, including saved results and a failure case.

Every AICR page

Page	Use it to
The AICR entries in the Catalog	Find the recipe for your service, accelerator and workload.
One AICR v1.0.0 recipe	Read one recipe end to end. It is published as OCI and stored in ConfigHub.
One AICR v0.20.0 recipe	See a recipe promoted through development, staging and production.
Questions AICR users ask	Get answers with the commands that give them.
Compare two GPU nodes	See how two nodes differ, with no recipe.
One recipe under Argo CD	Read what happened when one recipe ran on a throwaway cluster.
Where an AICR configuration comes from	Check which NVIDIA recipe an entry came from.
What the AICR entries refuse	See what the Catalog will not claim for AICR.

Choose the question first

New to AICR, with no GPU or cluster? Start with the local configuration exercise. You will inspect retained files and their identity; you will not deploy them.

Already operate GPU nodes? Compare node snapshots. This needs cluster access and permission to create the temporary collector resources described above.

How inspection, destination checks and runtime proof differ

Question	AICR path
What do I have?	Use snapshot and diff to report differences between existing GPU nodes. No recipe, bundle, or Catalog match is required.
What will it produce?	Select a provider-curated AICR leaf variant and inspect the exact files it generates.
Can this destination accept it?	Check that the destination matches the variant's GPU, network, cloud, controller, credential, API, and component requirements.
Did it work?	Run recipe-dependent resource and runtime checks only after the declared components have been deployed.

Path A: compare existing GPU nodes

Needs an existing Kubernetes cluster and collector permissions. The snapshot creates a temporary Job, ServiceAccount and RBAC; it does not change the node configuration. Use AICR's observation path when the question is whether two nodes or two points in time differ. It records kernel command-line settings, modules, system services, GPU hardware, and other measured state as YAML.

aicr snapshot --output baseline.yaml
# select the other node or repeat after the change
aicr snapshot --output current.yaml
aicr diff --baseline baseline.yaml --target current.yaml --fail-on-drift

A difference is not automatically a fault. Missing iommu=pt or nvidia_peermem is an observation first. A node without Mellanox networking may correctly omit both settings; a recipe intended for RDMA may require them.

Choose the recipe the provider intends for that node's service, accelerator, operating system, workload intent, platform, and relevant hardware before deciding what should change. NVIDIA curates the built-in AICR recipes. Other catalog providers can publish and review additional recipes.

Same observed target	Result
Assigned to standard networking	Pass. The two RDMA settings do not apply.
Assigned to Mellanox RDMA	Two findings: iommu=pt and nvidia_peermem.

The complete snapshot walkthrough includes a working local review command, a machine-readable result with both snapshot and profile hashes, optional local OCI output, and the commands that retain the review as non-deployable ConfigHub Units.

expected-resources answers a later question. It needs the selected variant and its declared components deployed. If those components are absent, record that check as blocked or not run; do not call it failed GPU conformance.

Read the maintained result · Open the complete review YAML · Open the AICR v0.20.0 release · See the tested assessment boundaries

Path B: inspect a retained configuration

This path reads a reviewed AICR-generated package. It does not inspect a live GPU node or prove that the selected platform runs.

Word	What it means here
Recipe	The AICR recipe that NVIDIA curates, selected before generation.
Base	The exact generated objects, digest, requirements, and evidence kept by the Catalog or ConfigHub.
Variant	A later environment or policy change linked to that base.

Which AICR configuration is which

The Catalog holds the same NVIDIA AICR recipe at several versions, and this page runs a smaller selection from one of them. Choose by what you want to do.

You want to	Use	What it is
Run something now with no GPU, cluster or account	The CPU starter on this page	Seven of the 17 Applications from NVIDIA AICR v0.14.0, selected by the Workshop because they need no GPU. It is not an upstream NVIDIA recipe. Read how the selection was made.
Inspect the whole recipe and take it into ConfigHub	The v0.20.0 entry	All 17 Applications of the EKS H100 training recipe with Kubeflow, with the charts inside them rendered and the five ConfigHub steps. It is the most fully recorded version.
See the newest NVIDIA release	The v1.0.0 entry	The same recipe at NVIDIA AICR v1.0.0, as 17 Applications. It is published as OCI. The charts inside its Applications are not rendered here.

None of the three has run on a GPU. Each entry page says what was and was not run for it.

1. Install ORAS

ORAS reads and writes OCI packages without running a container.

oras version

If that command fails, use the ORAS installation instructions.

2. Pull and check the configuration

Run one script. It uses an empty credential store for the public pull, checks the source digest, selects the seven reviewed files, and compares the local OCI with those files.

bash <(curl -fsSL https://confighub.github.io/helm-expt/site/sh/aicr-cpu-starter/try.sh)

Source: oci://europe-west1-docker.pkg.dev/nth-fort-499605-q5/helm-expt/aicr-eks-h100-training-kubeflow-argocd-config:0.14.0

Recorded source digest: sha256:dcf7feeeeaece04cb5d55cbc1106862172b3ae77718154252b39db1ad8957010

3. Read what you received

The result contains seven Argo CD Application files, the source-and-intent record that explains the selection, and a local OCI containing the same seven files.

find ./aicr-cpu-starter/config/templates -maxdepth 1 -type f -print
cat ./aicr-cpu-starter/source-and-intent.yaml
oras manifest fetch --oci-layout ./aicr-cpu-starter/aicr-cpu-starter.oci:0.14.0

Selected Application	Argo CD sync wave
`cert-manager`	2
`nfd`	3
`prometheus-operator-crds`	6
`kube-prometheus-stack`	7
`k8s-ephemeral-storage-metrics`	9
`kai-scheduler`	10
`prometheus-adapter`	15

Expected local OCI digest: sha256:62eaf39703ca0e5e968fcd4a667cbccbb5b3ff16403471c4b3441b3681b2b225.

Where an AICR configuration comes from records the provider, catalog and selected recipe behind the retained configuration, and what this example proves.

Choose what to do next

Keep the files and OCI locally, or upload it into ConfigHub when your team needs shared changes, environment variants, approvals, and promotion from development to production. That account step is the same for every configuration.

To gate and move a change to this AI-platform configuration through environments, compare the exact object sets and promote the one that passed.

For deployment, choose the controller or direct path that will consume the reviewed objects. Do not apply this platform configuration until you have reviewed its component requirements and changed the recorded storage-class residue.

Compare this with a native platform built from tested parts. Build a Kubara platform composes similar components without Argo CD Applications from AICR. Apps on a platform defines how a workload lands on either kind.

Compare the other inference examples · Try the shorter Redis example