Quote

Apex quick-start guide

8-GPU NVIDIA HGX systems, built and tested by the platform manufacturer. From unboxing to a verified first boot.

Applies to: Apex B300, Apex B200

Safety first. Never open a power supply or work on mains wiring; use a licensed electrician. If you see coolant or smell burning, power off and unplug immediately. Your components’ manufacturer documentation takes precedence over this general guide.

01Unbox and inspect

  1. Inspect the shipping carton before you open it. If it is crushed, punctured or wet, photograph it first.
  2. Unpack on a clean, dry surface and keep all packaging. You may need it to ship a part back under the manufacturer’s warranty.
  3. Check the system for visible damage: dents, loose parts rattling inside, or bent connectors.
  4. Report shipping damage to [email protected] right away, with photos. Do not power on a system that looks damaged.
  5. Let the system reach room temperature before you power it on if it arrived cold, so no condensation forms inside.

02Site requirements

  1. HGX B200 and B300 systems need data-center power and cooling: well over 10 kW per server. Agree power, cooling, floor loading and rack space with your facility before delivery.
  2. The chassis is very heavy. Use a server lift and follow the platform manufacturer’s installation guide for rails and handling.
  3. Plan the network: management (BMC), data network, and high-speed GPU networking if you connect several servers.

03Power

  1. Connect every power supply as the platform manufacturer specifies, across separate PDUs and circuits.
  2. Plug each power supply into its own outlet or PDU socket, not a shared power strip. Never use extension cords for a GPU system.
  3. If you are not sure a circuit can carry the load, have a licensed electrician check it. Don’t modify circuits, outlets or breakers yourself, and never open a power supply.

04Management controller (BMC)

  1. Connect the BMC port to your management network, log in to its web interface and change the default password straight away.
  2. Use the remote console to power on and watch the first boot. Keep the BMC off public networks.

05First boot and software

  1. Your system ships with the operating system and software you chose on your quote, for example Ubuntu with the NVIDIA driver, CUDA, PyTorch and vLLM, or Windows.
  2. On first boot, create your user account, connect to the network and install updates: sudo apt update && sudo apt upgrade. Reboot afterwards.
  3. Kernel updates can need a matching NVIDIA driver. If nvidia-smi stops working after an update, reinstall the recommended driver with sudo ubuntu-drivers install and reboot.
  4. Keep work in Python virtual environments or containers so system packages and the driver stay stable.

06HGX checks

  1. HGX systems connect their GPUs through NVSwitch. The nvidia-fabricmanager service must be installed, matching the driver version exactly, and running: check it with systemctl status nvidia-fabricmanager.
  2. Run nvidia-smi topo -m and confirm all eight GPUs connect to each other over NVLink.
  3. Run the NCCL tests to confirm GPU-to-GPU bandwidth before production jobs.

07Verify the system

  1. Run nvidia-smi. Every GPU you ordered should be listed with its full memory size and no errors.
  2. Check the CPU and memory with lscpu and free -h, and storage with lsblk.
  3. Confirm PyTorch sees the GPUs: python3 -c "import torch; print(torch.cuda.is_available(), torch.cuda.device_count())".
  4. Run a 15–30 minute load test of your real workload while you watch temperatures with nvidia-smi -l 5. GPUs should hold steady well below their throttle point.
  5. Write down the serial numbers of the system and its GPUs (nvidia-smi -q | grep -i serial). You need them for any warranty claim.

08Getting help

  1. For setup, configuration, performance and troubleshooting questions, use the AI technical assistant in the chat (bottom right of any page). Paste exact error messages and command output for the best answers. It is best-effort, so verify before acting.
  2. Hardware faults (a dead part, failing memory, a GPU missing after basic checks) are covered by the component manufacturer’s warranty. The assistant can help you identify the failing part and its serial number.
  3. Warranty terms: see our quote and order terms at /terms/.