Infrastructure Engineer

Overview

As a Principal Infrastructure Engineer, AI Cluster Performance & Validation, you will be a critical member of the AI Infrastructure Operations team, responsible for ensuring the acceptance, performance, and scalability of our cutting-edge AI and High-Performance Computing (HPC) environments. Leveraging software engineering and testing principles, you will focus on building and maintaining the control plane, tooling, and automation that supports performance and validation testing of large-scale AI clusters. Your work will directly translate into higher system availability, compute optimization and reduced operational costs.

Key Responsibilities

• Own the technical definition of "healthy at scale." Set the architecture, roadmap, and acceptance criteria by which multi-thousand-GPU clusters are declared production-ready, and establish the performance bar (collective bandwidth, job goodput, model FLOPs utilization) that every cluster must clear before and after customer handover. Establish technology and product direction in collaboration with other tech leads, managers, and senior leadership.
• Run and instrument real AI workloads as a diagnostic instrument. Stand up and execute distributed training and inference jobs — open-source and customer-representative models — across thousands of accelerators to validate cluster behavior under genuine load rather than synthetic proxies alone, and translate what those runs reveal into fleet-wide fixes.
• Lead deep diagnosis of large-scale cluster failures and performance regressions, isolating root cause across the full stack: GPU and NIC firmware, PCIe/NVLink topology and NUMA placement, InfiniBand/RoCE fabric health, congestion control and routing, storage and data-loader throughput, scheduler placement, and framework/communication-library behavior. Serve as the final escalation point for the hardest slow-job and stalled-job investigations.
• Design and build the validation and burn-in systems that qualify nodes, racks, and full pods at scale — NCCL/RCCL collective sweeps, HPL/HPCG and MLPerf-style benchmarks, thermal and power soak tests, straggler and flapping-link detection — and automate them so that qualification is a repeatable pipeline, not a manual campaign.
• Drive cluster optimization end to end, tuning fabric configuration (adaptive routing, QoS and congestion control, SHARP in-network reduction, rail and topology-aware placement), collective communication libraries and algorithm selection, GPUDirect RDMA and storage paths, and host-level settings (huge pages, IRQ affinity, CPU governors, MIG and driver configuration) to convert raw hardware into delivered throughput.
• to translate operational and customer performance needs into durable engineering solutions, and to feed diagnostic signal back into provisioning, remediation, and capacity workflows.