On November 2, NVIDIA reported preview Kubernetes recipes tested on early Vera Rubin NVL72 reference clusters.
The big change
- Versioned recipes help teams reproduce a known cluster configuration.
- Passing validation does not mean the recipe is production-qualified.
- Operators must check the recorded revision before applying validation to changed recipes.
Why a recipe matters
NVIDIA AI Cluster Runtime (AICR) packages compatible driver, operator and operating-system choices as versioned recipes. Its earlier v1.0 announcement described generating deployment bundles and checking a running cluster against the chosen configuration. This gives teams a reproducible starting point, while their own environment still needs verification.
BIG CHANGE previously covered cluster training recovery.
The tested stack includes an R615.23 driver, Ubuntu 26.04 and DRA driver v0.5.0. Training and inference configurations passed NVIDIA's deployment and conformance checks. The recipes lack supported-release lifecycle qualification. An unmapped Rubin architecture caused DRA plugin crashes during early validation; NVIDIA says driver v0.5.0 includes the fix.
What the test covers
Kubernetes dynamic resource allocation uses claims to request devices and schedules a workload where the requested resources are available. A driver and node components then configure device access. That separation matters when a distributed workload needs a consistent set of resources rather than simply counting GPUs.
NVIDIA reports a cross-node NVLink all-reduce test using ComputeDomains. This is not a training benchmark or customer deployment. Validation excluded multi-instance GPU profiles and ConnectX-9 RDMA performance.
NVIDIA says its then-current recipe would mismatch the signed historical digest. Operators must verify software, network and workload; production support or universal performance remains unproven.
