Skip to main content
Configure networking for GPU cluster nodes using public and private networks, routers, floating IPs, and security groups.

​
Network interfaces on a cluster node

Each node in a GPU cluster can have multiple network interfaces attached simultaneously: Private networks use VxLAN on Virtual GPU clusters and VLAN on Bare Metal GPU clusters.

​
Connect cluster nodes to a private network

A private network with a subnet isolates cluster nodes and provides a fixed IP address range. Without a router attached to the subnet, nodes cannot initiate outbound connections — attach one when nodes need to reach package repositories, object storage, or external APIs. For inbound SSH access, assign a floating IP to the head node and use a security group to allow SSH only from trusted IP addresses.

​
Use a public network

A public network interface assigns each node a direct public IP, so no router or floating IP is needed. Use this when all nodes must pull external resources independently — container images, datasets, or model weights on each node separately.

​
Connect VAST file shares

After creating a VAST file share, a dedicated network interface becomes available in the node’s network settings. Attach it to each node that needs to mount the share — this interface does not appear until the file share exists.

​
Cluster networking and GPU interconnect

Multi-node GPU communication uses InfiniBand on supported configurations and is set up automatically when the cluster is created. This interconnect is separate from the Cloud networking described above — it does not appear in network settings and requires no manual configuration.