We are hiring a Senior Platform Engineer to build, run and harden the bare-metal GPU platform behind our AI cloud. This is a hands-on Linux systems role at the intersection of HPC, GPU virtualisation and high-performance fabrics: you will own the servers, the OS, the GPUs, the DPUs and the InfiniBand fabric they run on.
We are a European Neocloud operating a large NVIDIA GPU estate (H100, H200 and B300) across multiple datacentres in Europe. Our customers run large-scale AI training and inference workloads, on both physical and virtualised GPU clusters, and they depend on our platform being fast, stable and predictable.
The platform you will own:
Compute:
Virtualisation: KVM/QEMU via libvirt on selected hosts; KubeVirt for VM workloads on Kubernetes; GPU passthrough and SR-IOV Virtual Functions.
DPUs: NVIDIA BlueField-2 and BlueField-3, operated in DPU mode.
GPU Fabric: InfiniBand on ConnectX-7 and ConnectX-8 adapters, managed with OpenSM and NVIDIA UFM.
Ethernet & Storage: Multi-tier spine-leaf Ethernet carrying block and object (blob) storage traffic; IPv6-first addressing.
Storage: NVMe-oF, RoCE and NFS; VAST Data and Pure Storage platforms.
Orchestration: Kubernetes with Operators, CNIs and CSIs; Docker / containers.
Tooling: Ansible (IaC), Nautobot, OpenBao, Defguard, PXE / preseed / cloud-init, Redfish / IPMI.
This is a platform engineering role, not a network engineering role. Networking, fabric and storage knowledge matter because they are part of the server platform you operate end to end.
YOUR RESPONSIBILITIES
- Build, tune and maintain Linux bare-metal GPU hosts for HPC and AI workloads, from BIOS/BMC through kernel, drivers and firmware.
- Deploy, configure and operate BlueField-2/3 DPUs in DPU mode, including BFB images, DOCA services, OVS offload and network/storage offload on the Arm cores.
- Operate and troubleshoot InfiniBand fabrics with OpenSM and UFM: subnet management, partitions (PKeys), routing, health monitoring, and performance analysis of NCCL and RDMA traffic.
- Deliver GPU virtualisation: PCIe passthrough, SR-IOV Virtual Functions on GPUs and ConnectX adapters, and integration of GPU VMs into Kubernetes via KubeVirt or libvirt/QEMU hypervisors.
- Build and support both physical and virtual GPU clusters interconnected over InfiniBand, down to NUMA, PCIe topology, GPUDirect RDMA and IOMMU level.
- Run IPv6-first host networking and support storage networks carrying NVMe-oF, RoCE, NFS, block and object traffic across the spine-leaf fabric.
- Automate provisioning and lifecycle with PXE, preseed, cloud-init, Redfish/IPMI and Ansible; review and extend our existing IaC code base.
- Write production-quality Bash and Python tooling for deployment, validation, health checks and fleet operations.
- Lead incident response for platform issues across multiple European datacentres, perform root-cause analysis, and drive permanent fixes.
- Work with networking, storage, Kubernetes and customer-facing teams to plan hardware generations (e.g. B300, ConnectX-8, BlueField-3) into production.
YOUR ESSENTIAL SKILLS (in priority order)
- Linux in an HPC configuration:
- Deep knowledge of the Linux OS and kernel: modules, drivers, sysctl/boot parameters, hugepages, NUMA, CPU pinning, IOMMU/VFIO, PCIe topology.
- Debian/Ubuntu in production; comfortable debugging at kernel level (dmesg, perf, ftrace, crash analysis).
- Complete understanding of NVIDIA GPUs and GPU concepts: drivers, CUDA, Fabric Manager, NVLink/NVSwitch, MIG, DCGM, GPUDirect RDMA.
- GPU virtualisation: passthrough, SR-IOV Virtual Functions, vGPU concepts, and VM integration with KVM/QEMU/libvirt or Kubernetes with the KubeVirt operator.
- NVIDIA BlueField in DPU mode (must have):
- Extensive hands-on experience with BlueField-2 and/or BlueField-3: provisioning, BFB images, firmware, DOCA, OVS/ASAP2 offload, and host-to-DPU networking.
- ConnectX-7 / ConnectX-8 configuration, firmware management (mlxconfig, mlxfwmanager) and SR-IOV.
- InfiniBand with OpenSM and UFM (must have):
- Operating IB fabrics with both OpenSM and UFM: subnet managers, partitions, routing engines, fabric diagnostics (ibdiagnet, perfquery, ibnetdiscover).
- Low-level RDMA concepts (verbs, QPs, GIDs, LIDs) and validating physical and virtual GPU clusters over IB (perftest, NCCL tests).
- Networking, with an emphasis on IPv6:
- Strong TCP/IP fundamentals for IPv6 and IPv4: addressing, SLAAC/DHCPv6, neighbour discovery, routing.
- Working knowledge of DHCP, DNS, NTP and BGP as they apply to hosts in a spine-leaf datacentre fabric.
- Storage networks:
- Block and object (blob) storage concepts over Ethernet: NVMe-oF (TCP/RDMA), RoCE, NFS, multipathing and performance tuning.
- Understanding how storage traffic behaves across a multi-tier spine-leaf network (QoS, PFC/ECN for RoCE, MTU).
YOUR EXPERIENCE
- Bash and Python: excellent Bash; excellent Python for automation and tooling (strong scripting and maintainable code, not necessarily software-developer level).
- Automation: working knowledge of Ansible to review and enhance an existing IaC code base.
- Server management and deployment: BMC, Redfish, IPMI; PXE, preseed, cloud-init.
- Containers: Docker/containerd and Kubernetes, including Operators, CNIs and CSIs.
- Incident handling: proven ability to lead incidents on production infrastructure, under pressure, through to root cause.
- Hands-on experience with VAST Data and/or Pure Storage.
- Immutable, Kubernetes-focused OS distributions such as Flatcar Container Linux or Fedora CoreOS.
- Operational experience with Nautobot, OpenBao (or HashiCorp Vault) and Defguard.
- General knowledge of HPC and AI concepts and tooling: Slurm, MPI, NCCL, distributed training and inference frameworks.
- NVIDIA GPU Operator and Network Operator on Kubernetes.
- Experience in a cloud provider, Neocloud or large HPC centre operating GPUs at scale across multiple sites.
About you:
- Typically 7+ years in Linux systems or infrastructure engineering, with recent hands-on time on GPU or HPC platforms.
- Calm and methodical during incidents; you dig to root cause rather than stopping at the workaround.
- You automate by default and leave behind documentation and tooling others can use.
- Clear communicator in English, comfortable working across teams and datacentres in different countries.