Skip to content

Requirements

Minimum Notes
Appliance 1 × NVIDIA DGX Spark or Asus GX10 (GB10, 128 GB unified memory) More boxes run larger models: see Add boxes
Disk 150 GiB free under /var/lib (200 GiB recommended) Models are 5–100 GiB each
Pair cabling (optional) 1 × QSFP56/QSFP28 direct-attach cable per pair Port 1 to port 1; no switch needed
  • DGX OS (Ubuntu 24.04 based) as shipped, with Docker and the compose plugin (both included).
  • git and curl (sudo apt install git curl if missing).
  • A user account with sudo. Do not install as root.
  • A fixed LAN address for the appliance (a DHCP reservation is fine).
  • Internet access during installation to download the software and the AI models (about 100 GiB on first install). After that the appliance works without internet, except for the cloud, SIEM and threat-intelligence services you connect.
  • Systems you enrol need to reach the appliance on TCP 8080 (outbound from the system).
Port Who connects What
443 Browsers The console (and Grafana under /grafana/)
8080 Agents and collectors Agent connections, installers, log and metric ingest
22 Administrators SSH, for installation and maintenance

Everything else (the database, the AI gateway, the models) listens only on the appliance itself or is protected by the appliance’s certificates. See Ports and firewall rules.

Collect these before you start connecting things (each is optional):

To connect You need
LevelBlue USM Anywhere Your USM Anywhere address and an API client ID and secret
Elastic Security The Elasticsearch URL and an API key with read access to the alert indices
AWS An access key with read access (add write permissions only for changes you want Garrison to make)
Azure A service principal: subscription, tenant, client ID and secret
Kubernetes A kubeconfig
FortiGate The management URL and a REST API admin token (read-only is enough to start)
Cloudflare WAF An API token with zone WAF read (and edit, for approved rules) and the zone ID
AWS WAF Access key, region, scope and the web ACL name