Senior Principal Architect - HPC & AI
Posted
Description
Job Overview
We design and deliver some of the largest and most demanding HPC and AI environments in the industry — GPU-dense AI factories measured in hundreds of accelerators, research clusters with decade-long lifespans, and inference estates that have to be commercially defensible as well as technically sound. The Principal Architect for HPC & AI owns the technical blueprint for that work across the organisation.
This is the most senior individual contributor role in the architecture practice. You are the design authority: the person whose judgement sets what we build, how it is validated, and what we refuse to build. You will own reference architectures and design standards, sit on architecture boards and design gates, and be the final technical escalation on the hardest engagements — while still writing the HLD for the deal that matters most.
The role is centred on the NVIDIA ecosystem, and that means being genuinely current rather than conference-current: DGX and HGX platforms, Blackwell-generation systems, NVLink domains, Spectrum-X and Quantum fabrics, and the reference architectures and validated designs that NVIDIA publishes. You will be our technical counterpart to NVIDIA's solution architects and engineering teams, translating their direction into designs that survive contact with real customer sites, real power budgets, and real operational teams.
Key responsibilities:
Architecture and design
We design and deliver some of the largest and most demanding HPC and AI environments in the industry — GPU-dense AI factories measured in hundreds of accelerators, research clusters with decade-long lifespans, and inference estates that have to be commercially defensible as well as technically sound. The Principal Architect for HPC & AI owns the technical blueprint for that work across the organisation.
This is the most senior individual contributor role in the architecture practice. You are the design authority: the person whose judgement sets what we build, how it is validated, and what we refuse to build. You will own reference architectures and design standards, sit on architecture boards and design gates, and be the final technical escalation on the hardest engagements — while still writing the HLD for the deal that matters most.
The role is centred on the NVIDIA ecosystem, and that means being genuinely current rather than conference-current: DGX and HGX platforms, Blackwell-generation systems, NVLink domains, Spectrum-X and Quantum fabrics, and the reference architectures and validated designs that NVIDIA publishes. You will be our technical counterpart to NVIDIA's solution architects and engineering teams, translating their direction into designs that survive contact with real customer sites, real power budgets, and real operational teams.
Key responsibilities:
Architecture and design
- Own the end-to-end architecture for HPC and AI factory environments: workload profiling and sizing model, compute and accelerator selection, fabric and storage topology, rack, power and cooling envelope, and the software and scheduler stack above it.
- Author reference architectures and validated designs, with the performance model documented: what the design delivers, at what scale, under which assumptions, and where it degrades.
- Own HLD/LLD, bills of materials, and the acceptance criteria for flagship engagements — personally where the stakes are highest, and by review and sign-off where the team is delivering.
- Translate customer AI ambitions into engineered, supportable platforms: model class and training versus inference mix, growth and refresh path, multi-tenancy boundaries, and the operational model.
- Chair architecture boards and design gates; produce architecture decision records, and ensure designs comply with vendor support matrices, security baselines, and our internal standards.
- Act as the organisation's senior technical authority on the NVIDIA stack: DGX and HGX systems, DGX SuperPOD, B200/B300 and GB200 NVL72-class platforms, NVLink and NVSwitch domains, Spectrum-X Ethernet and Quantum InfiniBand fabrics, Base Command Manager, and NVIDIA AI Enterprise.
- Adapt NVIDIA reference architectures and validated designs to real customer environments rather than dropping them in unchanged — including AI Factory and NAIC-aligned blueprints, and the power, cooling, and facilities constraints those blueprints assume.
- Own the technical relationship with NVIDIA solution architecture and engineering: design reviews, defect and interoperability escalations, early-access and roadmap engagement, and partner-programme technical alignment.
- Track the NVIDIA roadmap and translate announcements into delivery, cost, skills, and support implications for our portfolio — ahead of the customer asking.
- Build and govern the HPC and AI reference catalogue and build standards: firmware, BIOS, driver and CUDA stack pinning, cluster provisioning model, naming and topology conventions, and the health-check and drift-detection regime that keeps a running cluster inside its validated envelope.
- Define validation and acceptance for each platform — burn-in and stress, HPL/HPL-MxP and MLPerf baselines, NCCL collective benchmarks, fabric and storage validation, failure injection, and thermal and power-capping validation — and sign off platform readiness.
- Own the benchmarking and lab capability that underpins our claims: reproducible results, published method, and the discipline not to quote a number we cannot re-run.
- Feed delivery and operational learning back into standards, the portfolio roadmap, and vendor roadmap conversations.
- Lead design workshops, architecture reviews, and customer technical boards; drive open issues to technical closure and hold the design line under schedule pressure.
- Serve as final technical escalation for faults spanning the full stack: GPU XID, ECC and NVLink errors, thermal throttling and power capping, PCIe AER, firmware and driver interactions, fabric congestion and collective stalls, parallel filesystem and scheduler behaviour, and silent performance regression at scale.
- Support build and handover on strategic accounts: validate as-built versus as-designed, resolve drift, and ensure runbooks and operational documentation are good enough to keep the platform in-envelope without you.
- Provide organisation-level technical direction across the HPC and AI practice; mentor Compute and Network Architects and senior engineers, and set the technical bar for hiring.
- Lead pre-sales on strategic opportunities: solution shaping, RFx technical responses, sizing, cost and effort modelling, and credible technical presentations to customer executives.
- Represent the organisation externally — NVIDIA partner forums, conferences, reference customers, and published technical content — and grow our standing in the ecosystem.
- NVIDIA ecosystem depth, current and demonstrable — hands-on architecture experience across DGX and HGX systems and Blackwell-generation platforms, NVLink and NVSwitch topology, MIG and GPU partitioning, GPUDirect, and the tooling that surrounds them. You should be able to defend a design choice to an NVIDIA solution architect and to a customer's GPU performance team.
- HPC architecture and workload fluency — practical command of the layers around the accelerator: parallel filesystems (Lustre, IBM Storage Scale, WekaFS), MPI and collective communication, schedulers such as Slurm or PBS Pro, CUDA and maths libraries, and benchmarking with HPL/HPL-MxP and MLPerf. You understand why a cluster runs at 60% of peak and how to close the gap.
- AI and HPC fabric design — design-level expertise in InfiniBand (Quantum-generation, subnet management, SHARP) and/or Spectrum-X Ethernet, including rail-optimised and non-blocking topologies, RDMA and RoCE behaviour, and how collective traffic patterns interact with fabric design.
- Reference-architecture and AI Factory design — experience designing against NVIDIA reference architectures and validated designs, and adapting AI Factory or NAIC-aligned blueprints to real sites with real power, cooling, and facilities constraints.
- Design authority at scale — a record of owning HLD/LLD and bills of materials on large, high-value programmes, running design workshops and architecture boards, authoring standards, and defining and enforcing validation and acceptance criteria.
- Cross-domain troubleshooting — the ability to take an ambiguous symptom and drive it across compute, fabric, storage, power, cooling, and software to root cause, with vendors engaged rather than deferred to. Principal-level judgement about when to keep digging and when to change the design.
- Principal-level influence — credible with customer architects, CxO stakeholders, and vendor engineering in the same week; able to set technical direction, disagree with evidence, and bring a room to a decision.
- NVIDIA AI Enterprise and platform software — NIM microservices, NeMo, Base Command Manager, GPU Operator and Kubernetes scheduling for AI workloads, and Run:ai-style GPU orchestration.
- Liquid cooling and high-density mechanical design — direct-to-chip, CDU and rear-door heat exchangers, and the facilities interface for 100kW+ racks.
- Storage architecture for AI — parallel and scale-out filesystems, GPUDirect Storage, checkpoint and burst-buffer design, and data pipeline and curation tooling.
- Cloud and hybrid AI — DGX Cloud or CSP accelerator estates, burst-to-cloud and hybrid scheduling, and cost modelling for reserved versus on-demand capacity.
- Automation and cluster lifecycle — Ansible, Terraform, or Kubernetes-based provisioning; image lifecycle management; and monitoring with DCGM and fabric telemetry.
- Security, compliance, and sovereignty — designing for regulated and sovereign AI requirements, data-residency constraints, and accreditation regimes.
- Recognised standing — NVIDIA certifications (AI Infrastructure / NCP), published technical work, conference speaking, or equivalent peer recognition in the HPC and AI community.
- Open communication and a focus on psychological safety
- Recognition of achievements and contributions
- Collaboration across multicultural teams
- Less bureaucracy, more focus on results
- Technical Interview
- Managerial Interview
- Client Interview
- Pre-offer stage + Reference Check (if requested)
- Official Offer
Skills
Sign in to read and join the discussion.