Senior Network Production Engineer, AI Supercomputing

We are building and operating frontier-scale AI supercomputers used to train the world’s most advanced models. The backend network is a critical part of the machine: a single degraded NIC, switch, link, or path can reduce training performance or disrupt jobs spanning thousands of GPUs.

We are looking for a senior, hands-on Network Production Engineer to own the production health and operations of our multi-rail Ethernet backend network built on MRC and RDMA technologies. You will bridge network engineering, distributed systems, hardware health, and production operations to ensure the network delivers predictable performance at extreme scale.

Your mission is simple: make the network invisible to researchers by maximizing large-job success, minimizing performance degradation, and recovering quickly when failures occur.

Own Production Network Health

• Own the availability, performance, and operational readiness of the MRC Ethernet backend network.

• Define and operate service-level indicators for packet loss, congestion, link health, path diversity, bandwidth, tail latency, collective performance, and job impact.

• Detect slow or degraded components before they cause training failures or reduce model FLOPs utilization.

• Build health models that correlate switch, NIC, host, topology, and application telemetry.

Operate the Network at Frontier Scale

• Participate in on-call rotation and lead the response to high-severity network incidents.

• Diagnose failures spanning GPUs, NICs, switches, cables, firmware, drivers, network operating systems, MRC, NCCL, Kubernetes, and training workloads.

• Develop automated mitigation mechanisms, including path avoidance, node quarantine, workload relocation, and safe component remediation.

• Create operational procedures for maintenance, upgrades, rollback, capacity expansion, and topology changes.

Improve Large-Job Reliability and Performance

• Work directly with training teams to investigate collective-performance degradation, stalls, timeouts, job restarts, and unexplained MFU loss.

• Translate low-level network signals into clear job-level impact.

• Establish network qualification gates for admitting nodes and racks into large training pools.

• Run failure-injection and resilience testing to validate behavior under link, NIC, switch, and path failures.

• Improve placement and routing policies for jobs spanning racks, rows, and network planes.

Automate Fleet Operations

• Build production-quality software and automation for network validation, monitoring, diagnosis, remediation, and fleet-wide change management.

• Replace manual investigation with deterministic workflows and actionable alerts.

• Automate firmware, driver, configuration, and network operating-system compliance checks.

• Maintain authoritative network inventory, topology, and configuration state.

• Reduce mean time to detect, isolate, mitigate, and permanently resolve failures.

Drive Cross-Company Execution

• Lead technical investigations involving internal infrastructure teams, cloud and datacenter operators, Azure Networking, NVIDIA, switch and NIC teams, and network-software partners.

• Own incidents through resolution, even when the underlying failure crosses organizational boundaries.

• Produce clear root-cause analyses with corrective actions, owners, and completion dates.

• Convert recurring production failures into requirements for future network, server, and accelerator architectures.

• Create runbooks and train engineers across regions to provide reliable round-the-clock coverage.

What Success Looks Like

Within the first year, you will:

• Establish clear operational ownership and measurable service levels for the MRC backend network.

• Build end-to-end visibility from physical links and network paths to collective performance and training-job health.

• Materially reduce network-caused job interruptions, degraded-performance hours, and unresolved incidents.

• Reduce detection and isolation time from hours to minutes through automation.

• Safely qualify and release new racks and network capacity into production.

• Create a scalable operating model that supports increasingly large training jobs without proportional growth in manual operational effort.

Primary success metrics include:

• Large-scale training job success rate and goodput

• Network-attributed lost GPU hours

• Collective performance and tail latency

• Mean time to detect, isolate, mitigate, and resolve failures

• Percentage of failures detected before workload impact

• Change failure and rollback rates

• Automated remediation coverage

• Time from hardware availability to production-ready capacity

Minimum Qualifications

• Significant experience operating large-scale datacenter or high-performance computing networks in production.

• Deep understanding of Ethernet, RDMA, RoCEv2, congestion control, routing, load balancing, and lossless or near-lossless network design.

• Experience debugging complex failures across switches, NICs, hosts, drivers, firmware, and distributed applications.

• Strong Linux systems knowledge and proficiency in Python, Go, C++, or another systems-oriented programming language.

• Experience building monitoring, automation, diagnostic, or remediation systems for production infrastructure.

• Demonstrated ability to lead high-severity incidents and coordinate resolution across multiple engineering teams.

• Strong analytical skills and the ability to translate telemetry into a defensible root cause.

• Willingness to participate in a global 24/7 on-call rotation.

Preferred Qualifications

• Experience operating GPU training clusters or other tightly coupled distributed-computing systems.

• Knowledge of MRC or other multi-rail, multi-plane, multipath Ethernet transports.

• Experience with NCCL, collective communications, CUDA, GPUDirect RDMA, and distributed training frameworks.

• Familiarity with ConnectX-class NICs, modern Ethernet switch ASICs, SONiC, SAI, and switch telemetry.

• Experience with InfiniBand and the operational differences between InfiniBand and Ethernet-based AI fabrics.

• Familiarity with Kubernetes, Slurm, workload placement, and cluster-health certification.

• Experience with streaming telemetry, gNMI, Prometheus, Datadog, Kusto, eBPF, or equivalent observability systems.

• Experience managing fleet-wide firmware, driver, or network operating-system changes.

• A track record of improving reliability through automation rather than recurring manual intervention.

Attributes We Value

• You treat the network as part of the training system, not as an isolated infrastructure layer.

• You are equally comfortable analyzing packet-level behavior, reading code, and coordinating a production incident.

• You stay accountable until the user-visible problem is resolved.

• You distinguish symptoms from root causes and measurements from assumptions.

• You automate recurring operational work and eliminate entire classes of failure.

• You communicate clearly under pressure and establish unambiguous ownership.

• You optimize for usable training capacity, reliable job execution, and researcher velocity.