Skip to main content

Private Virtual Cloud (VPC)

NovaPod VPC enables you to deploy GPU instances, storage volumes, and serverless worker pools within a cryptographically isolated private network. Traffic between your instances traverses encrypted internal wireguard mesh tunnels without ever hitting the public internet.

Key Benefits​

  • Zero Public Exposure: Keep database nodes, Ray head nodes, and distributed training all-reduce workers completely off public IPv4/IPv6 address spaces.
  • Microsecond Latency: Internal datacenter interconnects provide low-latency networking ideal for PyTorch DDP, DeepSpeed, and Megatron-LM.
  • Custom CIDR Subnets: Define your own private IP ranges (e.g., 10.100.0.0/16) to easily bridge into existing enterprise VPNs or hybrid cloud infrastructures.

Creating a VPC​

Via CLI​

# Create a private VPC in US-East
np vpc create --name training-cluster-vpc --cidr 10.10.0.0/16 --region us-east

# Launch a GPU instance inside the VPC
np launch \
  --gpu H100_SXM \
  --count 4 \
  --vpc training-cluster-vpc \
  --private-ip 10.10.1.10 \
  --name rank-0-master

Inter-Node Communication​

Instances attached to the same VPC can communicate directly using their private IP addresses:

# Test internal private connectivity from Worker 1 to Master
ping 10.10.1.10
nc -zv 10.10.1.10 29500  # Default PyTorch Master Port