Out-of-band console access for AI infrastructure and GPU clusters

Details

How do you reach every BMC in a GB300-class rack?

Cable each BMC, compute node and fabric switch console to one SLC 9000 and the whole rack sits behind a single console path. Units come in 16, 24, 32 and 48 ports in one rack unit, so count the consoles in the rack design and pick the size. Port counts are factory-fit, so size for the rack you’re building, not the one you have today.

That path doesn’t ride on the fabric. The SLC 9000 has its own two 1GbE and two 1/10G SFP+ uplinks with failover, and every SKU ships with dual power supplies, AC or DC.

Why stage a GPU rack over the console first?

Because the console works before the fabric does. You can watch every node and BMC boot, run your own baseline scripts over SSH on TCP 3000 plus the port, and check the output, all before the rack’s first production packet. Problems found here are cheap to fix. The same path stays in place for the life of the rack, including the day something does go wrong.

Engineers get there two ways. Percepxion opens a browser console to any attached device with no VPN and without knowing the device’s IP. For scripted access, each port answers direct SSH on TCP 3000 plus the port number, so port 12 is TCP 3012.

How does it fit into cluster automation?

The same tooling that builds the cluster can drive the console layer. The lantronix.oob collection, version 1.2.0 with 22 modules, is certified and published on Red Hat Automation Hub. Lantronix also publishes two MCP servers: Percepxion, with 38 tools for fleet-scale work, and SLC, with 33 tools against a single unit. Once ports are defined as managed devices, a playbook or an AI agent addresses them by name instead of by a port map. The console layer becomes one more target in the pipeline, not a separate system someone checks by hand.

What does the appliance expose on the management network?

Very little. SLC 9000 firmware exposes no shell, and containers you run for telemetry exporters are sandboxed at 256 MB of memory and one CPU core, so they can’t starve console traffic. That’s room for a node-exporter or Telegraf instance next to the rack it watches. The unit registers to Percepxion over an outbound-only MQTT connection, which means no inbound firewall rule into the management network.