NVIDIA Network Operator v26.4.0

Host Device

Direct passthrough to host network devices, with no virtualization layer. Each pod gets exclusive access to a physical NIC. Minimal CPU overhead — best for legacy HPC codes, specialized protocols, and DPDK applications. Works with both Ethernet and InfiniBand fabrics.

  • Network Operator Helm chart installed (see Installation).

  • l8k binary on your PATH.

  • One or more Mellanox NICs per node not used by the host.

Copy
Copied!
            

l8k discover --kubeconfig ~/.kube/config \ --save-cluster-config ./cluster-config.yaml

Copy
Copied!
            

l8k generate --user-config ./cluster-config.yaml \ --deployment-type host_device --multirail \ --save-deployment-files ./deployments

Add --fabric ethernet or --fabric infiniband if you need to constrain the fabric explicitly.

Copy
Copied!
            

l8k deploy

Copy
Copied!
            

kubectl -n nvidia-network-operator get pods kubectl get hostdevicenetwork

Previous SR-IOV InfiniBand
Next IP over InfiniBand (RDMA Shared)
© Copyright 2025-2026, NVIDIA. Last updated on Jun 14, 2026