> Markdown version of [/jobs/ext/4785279-nvidia-gpu-infrastructure-deployment-engineer-dgx-hgx-infiniband-ai-clusters](https://www.wearedevelopers.com/jobs/ext/4785279-nvidia-gpu-infrastructure-deployment-engineer-dgx-hgx-infiniband-ai-clusters). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # NVIDIA GPU Infrastructure Deployment Engineer (DGX/HGX, InfiniBand, AI Clusters) - **Company:** NVIDIA Ltd. - **Location:** United States - **Experience:** Experienced - **Contract:** Contract - **Skills:** Artificial Intelligence, Airflow, Data Centers, Linux, Ethernet, Network Interface Controllers, Firmware, InfiniBand, Network Control, Node.Js, AI Infrastructure, Scripting, Graphics Processing Unit (GPU), System Availability, Slurm, Hardware Infrastructure - **Published:** November 19, 2026 - **Apply:** https://www.dice.com/job-detail/86713b80-d3fb-4327-b192-826ee2b8c906 ## About the Role * 5-8 years in data centre infrastructure engineering, of which at least 2-3 years hands-on with NVIDIA GPU hardware. This matches NVIDIA's own stated prerequisite for its AI infrastructure certification. * Delivered at least one multi-node GPU cluster build. Node count, GPU model, fabric type and the candidate's own scope must be stated, not implied. * Comfortable on a customer site, under programme pressure, working to someone else's design and method of procedure. * Linux systems fluency; scripting for repeatable build steps. Certifications - strongly preferred: * NVIDIA-Certified Professional: AI Infrastructure (NCP-AII) * NVIDIA-Certified Professional: AI Networking (NCP-AIN) for fabric-leaning candidates ## Description Builds what the architect designed and proves it works. Hands-on from rack and stack through to signed-off acceptance: physical installation, firmware and out-of-band configuration, fabric bring-up, cluster software install, and the validation and burn-in that demonstrates the cluster performs to specification. This is a build-and-prove role. The validation half matters as much as the installation half, because an OEM-branded GPU build does not come with a ready-made acceptance suite - NVIDIA certification validates server models and configurations, not the cluster standing on a customer floor. The engineer runs and, where needed, assembles that acceptance testing., Physical build * Install GPU servers and GPUs; verify seating, airflow and serviceability. * Identify and validate cables and transceivers across the compute, storage and management fabrics; confirm signal quality before handover. * Confirm power and cooling parameters against the design, and escalate where the data hall will not support it. Bring-up and configuration * Configure BMC and out-of-band management; set up TPM where required. * Apply and verify firmware across GPU baseboards, NICs, switches, DPUs and transceivers; interpret and act on fault detection output. * Install and configure the control plane: NVIDIA Base Command Manager including high-availability head nodes, operating system and cluster components, Slurm with Enroot and Pyxis, GPU and DOCA drivers, container toolkit and NCLI. * Set initial parameters for third-party storage and confirm throughput against the design. Fabric * InfiniBand: provisioning, high availability, partition keys for multi-tenancy, quality of service, adaptive routing, and link and bandwidth monitoring through UFM. * Or Spectrum-X Ethernet: RoCE, ECN and PFC configuration, quality of service, adaptive routing, telemetry. * Troubleshoot with the standard tooling - ibdiagnet, iblinkinfo, ib_write_lat and ib_write_bw, ibping, What Just Happened, UFM health. Validation and acceptance - the half that is usually missing * Single-node stress and burn-in. * HPL and single-node NCCL, including NVLink and NVLink Switch verification. * ClusterKit node assessment to find the weak node before the customer does. * NCCL east-west fabric bandwidth tests across the cluster. * Sustained burn-in using NCCL, HPL and NeMo workloads. * Storage throughput testing under realistic checkpoint load. * Produce the acceptance report: what was tested, what it returned, what was remediated. Operate and remediate * Diagnose and replace failed GPUs, network cards, fans and power supplies. * Tune storage and fabric where results fall short of specification. * Hand over to the customer or operations team with documentation that matches what was actually built.