GB300 AND INFINIBAND (TM70ST RM 4598)

Position: GB300 AND INFINIBAND (TM70ST RM 4598)

Job Description
Responsible for the physical build, bring-up, validation, and fault resolution of GPU compute and its InfiniBand fabric, from rack-level installation through to documented test sign-off. Reports to a Principal Architect who owns the design and any required adjustments to the Low Level Design.

Bring-up and configuration

• Apply the approved firmware, driver, and operating system baseline across nodes, network interface cards, and switches.
• Bring up the InfiniBand fabric, including subnet manager configuration, topology verification, and link-state validation.
• Confirm that the as-built fabric matches the design, covering cabling, port mapping, speeds, and topology.

Testing and validation

• Run the agreed acceptance tests, including link and bandwidth tests, NCCL collectives, burn-in, and stress runs, recording results against pass and fail criteria.
• Identify and isolate faults such as link flaps, high error counters, failed optics or cables, and node or GPU failures.

Fault resolution and escalation

• Perform hands-on fixes, including reseating, swapping, re-cabling, and reflashing.
• Log issues with clear evidence and escalate to the vendor or to internal leads where outside
  scope.

Documentation and reporting

• Maintain accurate as-built records, test logs, and issue trackers.
• Provide daily progress and blocker updates to the delivery lead.
• Follow site safety, change control, and access procedures at all times.

Required skills

• Hands-on experience with GPU server and rack-scale deployments, including liquid-cooled systems.
• Working knowledge of InfiniBand, covering subnet management, fabric diagnostics, optics, and cabling.
• Linux administration and scripting for automation and test execution.
• Comfortable with firmware and driver lifecycle management.
• Clear written communication for logs, handovers, and escalations.

*************************************************************************************************************