​
Instance Infrastructure

​
Hardware Specifications

  • CPU: E series CPU instances (Azure)
  • GPU Options:
    • A100 80GB (default for both modes)
    • T4 16GB (available via CLI)
    • A100 40GB (available via CLI)
  • Memory: 24 vCPUs and 220 GiB RAM per GPU (Production mode, Prototyping is fully customizeable)
  • Location: United States (region varies)

​
Pre-installed Software

  • CUDA: Version 12.9
  • CUDNN: Version 9.0
  • PyTorch: Version 2.7.1
  • JupyterLab: Pre-installed
  • Docker: Available (see “Docker on Thunder Compute” guide)
  • Additional scientific Python libraries (NumPy, Pandas, etc.)
Do not attempt to reinstall CUDA. If compatibility issues arise, upgrade your other dependencies (e.g., PyTorch) rather than downgrading CUDA.

​
Networking

  • Egress/Ingress: 7 Gbps
  • IP Address: Dynamic
  • Firewall: All ports open by default

​
Port Access

  • CLI: Use tnr connect <instance_id> -t <port> to tunnel ports
  • VS Code: Use the built-in port forwarding feature
  • External access: Use Cloudflared: cloudflared tunnel --url http://localhost:<port>

​
Storage

​
Persistent Disk

  • All instances use persistent disks by default
  • Data persists when instances are stopped (storage charges continue)
  • Performance: ~100k IOPS and 1200 Mbps Read/Write

​
Snapshots

Save instance state for long-term storage at ~75% lower cost. See our Using Snapshots guide for details.

​
Data Retention

  • Inactive accounts: All storage deleted after 60 days of no running instances
  • Scope: Per account (not per instance)
  • Reset: Starting any instance resets the 60-day timer
  • Permanent: Deletions cannot be undone