01Overview
Job Purpose
Build and operate large scale GPU compute pods to deliver predictable, high throughput, low latency training and inference services across 10K GPU cluster.
Roles & Responsibilities
- Implementation
Stand up multi pod GPU clusters (rack/power/cooling layouts; TOR/leaf connectivity; IB/Ethernet host configs).
Implement GPU partitioning (MIG/vGPU profiles) and quota policies for multi tenant environments.
Integrate cluster schedulers (Slurm/Kubernetes) with GPU device plugins, node feature discovery, and accounting/quotas.
- Operations
Own day 2 operations across firmware/driver/DCGM/NVML lifecycles; execute change windows with zero/low downtime.
Capacity planning (GPU/CPU/Memory/NIC) and bin packing strategies for heterogeneous GPU SKUs.
- Performance & Optimization
Tune NCCL/UCX, GPU clocks/persistence, GPU Direct Storage, NUMA/locality, and CUDA runtime parameters.
Drive benchmarking and acceptance (HPL, HPL AI, MLPerf like internal suites); track perf regressions with SLOs.
- Reliability & Incident
Lead P0/P1 incident response for GPU, CUDA, or scheduler issues; perform root cause and preventative actions.
Act as the primary technical lead during production outages, coordinating cross-functional teams across Networking, GPU Operations, Platform Engineering, Storage, and Application teams to restore services within SLA targets.
Perform detailed Root Cause Analysis (RCA) for network, GPU, CUDA, NCCL, RDMA, and scheduler-related failures, identifying underlying causes and implementing preventive and corrective actions.
Collaborate with platform and GPU engineering teams to resolve issues impacting CUDA jobs, Kubernetes scheduling, Slurm workload management, GPU resource allocation, and large-scale AI training environments.
Define golden images; implement node remediation (cordon/drain/reimage) and auto healing workflows.
- Security & Compliance
Enforce GPU tenancy isolation (MIG, cgroup/device cgroup, mpsd), secure drivers/containers, SBOM an .
