AI Infrastructure Domain Architect
- Discuss this with your agent
- Open in Claude
- Open in ChatGPT
Prepare application
- Draft this with your agent
- Open in Claude
- Open in ChatGPT
Role details
Tech stack
+2 more
Job description
Designs accelerated-compute infrastructure for enterprise AI build-outs: GPU compute, east-west fabric, storage and the platform software that sits on top. This is a single-domain architect role, not a multi-discipline enterprise architect. Depth in NVIDIA-based infrastructure matters more than breadth across unrelated technology towers. The architect owns the solution from requirements through to a design an engineering team can build, and remains the technical authority during deployment. They are the one technical voice to the customer on the AI infrastructure stack and the escalation point for the build team. Responsibilities *
- Translate workload requirements - training, fine-tuning, inference, mixed - into a sized GPU cluster design covering compute, fabric, storage, power and cooling.
- Produce the design artefacts: high- and low-level design, bill of materials, rack and floor layouts, power and thermal budgets, cable plans, IP and fabric addressing schemes.
- Select and justify platform: NVIDIA-branded systems versus an NVIDIA-partnered OEM build, with the trade-offs stated in writing.
- Design the east-west fabric and defend the choice between InfiniBand and Spectrum-X Ethernet for the given workload.
- Design the storage tier against real throughput and checkpointing requirements rather than capacity alone.
- Define the acceptance criteria and test plan the deployment team will execute, and sign off the results.
- Write the methods of procedure the engineering team works to, and review their deployment documentation.
- Act as technical escalation during build; resolve design issues without stopping the programme.
- Work alongside the customerβs and the integratorβs own architects, including where the integratorβs team is being skilled up in parallel., Experience on another HPC stack is additive, not a substitute. Candidates from IBM, Cray or similar HPC backgrounds are in scope provided NVIDIA GPU infrastructure is also in their history. Reference architectures: NVIDIA DGX BasePOD and DGX SuperPOD, or the NVIDIA Enterprise Reference Architectures. Must be able to explain where a design departs from the reference and why. Fabric: NVIDIA Quantum InfiniBand and/or Spectrum-X Ethernet. RDMA, RoCE, rail-optimised topologies, congestion control. BlueField DPUs and DOCA, and SuperNIC-class adapters. Storage: High-throughput parallel storage for AI - any of DDN, WEKA, VAST, Pure or equivalent - sized against GB/s per GPU rather than terabytes. Platform software: NVIDIA Base Command Manager, DGX OS, GPU drivers and CUDA, NVIDIA GPU Operator, container toolkit and NGC. Working knowledge of Slurm and Kubernetes as scheduling layers, including when each is appropriate. Facilities literacy: Rack power density, liquid and air cooling implications, and the practical limits a data hall imposes on a design.
Requirements
- 8-12 years in enterprise infrastructure architecture, of which at least 3 years on GPU, HPC or accelerated-compute infrastructure.
- Demonstrable delivery of at least one multi-node GPU cluster from design through to acceptance. Node count, GPU model, fabric and the candidateβs specific role must be stated.
- Customer-facing experience: able to present and defend a design to a customerβs technical leadership.
- Strong technical writing. Design documents are a deliverable on these engagements, not a by-product.
Certifications - preferred, not mandatory: *
- NVIDIA-Certified Professional: AI Infrastructure (NCP-AII)
- NVIDIA-Certified Professional: AI Networking (NCP-AIN)
- NVIDIA DGX platform or InfiniBand training paths
Apply for this position
This job is hosted externally. Click below to view the full posting and apply.
Prepare application
- Draft this with your agent
- Open in Claude
- Open in ChatGPT
Good distractions
Loading talks and stories from around this roleβ¦