nvidia-datacenter-bringup

NVIDIA datacenter GPU bring-up on Ubuntu 24.04 requires a deterministic install order: applying iDRAC firmware prerequisites, installing DOCA-OFED before the driver, enrolling MOK for signed DKMS modules, and activating NVLSM and Fabric Manager.

The correct install order requires installing DOCA-OFED before the NVIDIA open-modules driver, followed by MOK enrollment for secure boot and firmware activation. This deterministic sequence ensures proper NVLink fabric stack configuration and host validation.

Yes, you can deploy NVIDIA HGX B300 or B200 GPUs in an air-gapped environment. The deployment uses three-tier DOCA and CUDA mirroring guidance alongside Sneakernet file:// repositories with GPG keys to install the full NVLSM and DOCA-OFED stack without internet connectivity.

Dell iDRAC firmware prerequisites are mandatory for Dell HGX B300/B200 datacenter bring-up. You must establish firmware v1.4.30 baseline and use iDRAC ExtendedReset for activation before proceeding with DOCA-OFED installation and GPU driver setup.

You validate NVIDIA NVLink Fabric Manager and GPU health by running a fleet health-check script. This script confirms that GPUs are visible, Fabric Manager and NVLSM are active, and the NVLink fabric status is fully operational across the datacenter host.

Yes, NVIDIA datacenter GPU bring-up supports collaboration with gpu-operator in pre-installed driver mode. This integration allows the host setup to work in synergy with gpu-operator, utilizing CUDA validator triage to confirm successful deployment and system health.