You will also be an integral part of achieving industry-leading end-to-end performance benchmarking publications: If MLPerf (Training & Inference), Working closely with NVIDIA (Megatron-LM, TensorRT-LLM & DGX cloud) and the open-source community (llm-d, vLLM and all popular ML frameworks) speak to you, come help us demonstrate CoreWeave's performance reliability leadership in the field.

Responsibilities

• Strategy & Leadership - Define the multi-year benchmarking strategy and roadmap; prioritize models/workloads (LLMs, diffusion, vision, speech) and hardware tiers. Build, lead, and mentor a high-performing team of performance engineers and data analysts. Establish governance for claims: documented methodologies, versioning, reproducibility, and audit trails.

• Perf Ownership - Lead end-to-end MLPerf Inference and Training submissions: workload selection, cluster planning, runbooks, audits, and result publication. Coordinate optimization tracks with NVIDIA (CUDA, cuDNN, TensorRT/TensorRT-LLM, Triton, NCCL) to hit competitive results; drive upstream fixes where needed.

• Internal Latency & Throughput Benchmarks - Design a Kubernetes-native, repeatable benchmarking service that exercises CoreWeave stacks across SUNK (Slurm on Kubernetes), Kueue, and Kubeflow pipelines. Measure and report p50/p95/p99 latency, jitter, tokens/s, time-to-first-token, cold-start/warm-start, and cost-per-token/request across models, precisions (BF16/FP8/FP4), batch sizes, and GPU types. Maintain a corpus of representative scenarios (streaming, batch, multi-tenant) and data sets; automate comparisons across software releases and hardware generations.

• Tooling & Automation - Build CI/CD pipelines and K8s controllers/operators to schedule benchmarks at scale; integrate with observability stacks (Prometheus, Grafana, OpenTelemetry) and results warehouses. Implement supply-chain integrity for benchmark artifacts (SBOMs, Cosign signatures).

• Cross-functional & Community - Partner with NVIDIA, key ISVs, and OSS projects (vLLM, Triton, KServe, PyTorch/DeepSpeed, ONNX Runtime) to co-develop optimizations and upstream improvements. Support Sales/SEs with authoritative numbers for RFPs and competitive evaluations; brief analysts and press with rigorous, defensible data.

Requirements

• 10+ years building distributed systems or HPC/cloud services, with deep expertise on large-scale ML training or similar high-performance workloads.

• Proven track record of architecting or building planet-scale data systems (e.g., telemetry platforms, observability stacks, cloud data warehouses, large-scale OLAP engines).

• Deep understanding of GPU performance (CUDA, NCCL, RDMA, NVLink/PCIe, memory bandwidth), model-server stacks (Triton, vLLM, TensorRT-LLM, TorchServe), and distributed training frameworks (PyTorch FSDP/DeepSpeed/Megatron-LM).

• Proficient with Kubernetes and ML control planes; familiarity with SUNK, Kueue, and Kubeflow in production environments.

• Excellent communicator able to interface with executives, customers, auditors, and OSS communities.

Nice to have

• Experience with time-series databases, log-structured merge trees (LSM), or custom storage engine development.

• Experience running MLPerf submissions (Inference and/or Training) or equivalent audited benchmarks at scale.

• Contributions to MLPerf, Triton, vLLM, PyTorch, KServe, or similar OSS projects.

• Experience benchmarking multi-region fleets and large clusters (thousands of GPUs).

• Publications/talks on ML performance, latency engineering, or large-scale benchmarking methodology.

XML job scraping automation by YubHub