🛠 Deploying a High-Performance Edge AI Node on Xerxes Nano
Objectives
This advanced tutorial aims to configure the Xerxes Nano board as an autonomous Edge computing node. We will leverage NVMe storage for AI models, use the SFP+ cage for a dedicated fiber link, and enable PoE power for 1U rack installation without separate power cabling.
Required Hardware
- Xerxes Nano board with Jetson Orin Nano Super or Turing RK1 module.
- NVMe M.2 SSD (2230 or 2280 form factor depending on internal adapter).
- SFP+ optical module (10G) and appropriate fiber cable.
- PoE+ compatible network switch.
- USB-C cable for initial configuration.
- Screwdriver and mounting screws for rack installation.
Steps
- Hardware Preparation: Insert the M.2 SSD into the slot located under the compute module. Secure the SFP+ optical module in the designated cage. Ensure the power selection jumper is set to "PoE".
- System Installation: Connect the board via USB-C and flash the OS (Ubuntu JetPack or standard Linux) onto the NVMe SSD. Once complete, disconnect USB.
- Network Configuration: Connect the fiber cable to the SFP+ port. On the system, identify the network interface (e.g.,
eth1) and configure a static IP or DHCP via Netplan or NetworkManager. Verify the link withip link show. - Rack Mounting: Install the board horizontally in a 1U chassis or vertically in 1.5U. Connect the fiber cable to the PoE switch.
- Storage Optimization: Mount the NVMe filesystem with
noatimeoptions to reduce wear. Configure a cache directory for AI models on this fast storage. - Thermal Management: Connect the fan to the PWM connector. Adjust the fan curve via
nvidia-smior system tools to maintain optimal temperature under load.
Verification
Run lsblk to confirm the NVMe SSD is detected and mounted. Test network bandwidth with iperf3 between the node and a remote server. Verify that the system boots correctly without USB power, solely via the PoE cable.
Further Reading
Explore using the OCuLink port to connect an external GPU if the integrated module's power is insufficient for your specific workloads.