🛠 Deploying an AI NAS with Xerxes Nano

Objectifs

This tutorial guides the installation of a Network Attached Storage (NAS) server with AI capabilities on the Xerxes Nano carrier board. We will leverage the SFP+ port for high-speed fiber connectivity, the M.2 NVMe slot for storage, and the OCuLink connector to add an external GPU if needed. The goal is to create a robust local infrastructure for processing large datasets.

Matériel nécessaire

• Xerxes Nano carrier board
• NVIDIA Jetson Orin Nano Super or Turing RK1 module
• NVMe M.2 Drive (Gen 3)
• Compatible 10G/25G SFP+ module and fiber optic cable
• PoE or USB-C power supply (selected via jumper)
• SD card reader for initial installation
• Ethernet cables for management

Étapes

1. Hardware Preparation: Insert the Jetson or Turing module into the SO-DIMM connector of the Xerxes Nano. Secure the NVMe drive in the M.2 slot. Plug the SFP+ module into the dedicated cage.
2. Power Supply: Set the power jumper to "PoE" if using a PoE switch, or "USB" otherwise. Connect power. The PWM fan must be connected to the dedicated connector to prevent overheating.
3. OS Installation: Download the NVIDIA JetPack or Ubuntu Server image. Write it to an SD card. Insert the SD card and boot the board.
4. Network Configuration: Once connected via SSH, configure the SFP+ interface. Identify the interface (e.g., enp1s0) with ip a. Assign a static IP:

sudo nmcli con mod enp1s0 ipv4.addresses 192.168.1.10/24
sudo nmcli con mod enp1s0 ipv4.gateway 192.168.1.1
sudo nmcli con mod enp1s0 ipv4.dns "8.8.8.8"
sudo nmcli con up enp1s0

1. Storage: Format the NVMe as ext4 and mount it:

sudo mkfs.ext4 /dev/nvme0n1
sudo mkdir /mnt/storage
sudo mount /dev/nvme0n1 /mnt/storage

Vérification

Verify the fiber link with ethtool enp1s0 (should show 10G/25G). Test storage speed with fio. Ensure the AI module is detected via nvidia-smi or lspci. Temperature should remain below 80°C thanks to PWM control.

Pour aller plus loin

Explore the open-source "Xerxes Home Lab" tool on macOS for sensor monitoring. You can add an external GPU via OCuLink to accelerate AI inference tasks on stored data.