Director , Perf and Benchmarking
3 days ago
Bellevue
Job Description CoreWeave is the AI Hyperscaler™, delivering a cloud platform of cutting edge services powering the next wave of AI. Our technology provides enterprises and leading AI labs with the most performant, efficient and resilient solutions for accelerated computing. Since 2017, CoreWeave has operated a growing footprint of data centers covering every region of the US and across Europe. CoreWeave was ranked as one of the TIME100 most influential companies of 2024. As the leader in the industry, we thrive in an environment where adaptability and resilience are key. Our culture offers career-defining opportunities for those who excel amid change and challenge. If you're someone who thrives in a dynamic environment, enjoys solving complex problems, and is eager to make a significant impact, CoreWeave is the place for you. Join us, and be part of a team solving some of the most exciting challenges in the industry. CoreWeave powers the creation and delivery of the intelligence that drives innovation. What You'll Do: CoreWeave is seeking a Director of Performance to own our end-to-end performance strategy across training and inference. You will lead a multi-disciplinary organization spanning performance tooling & debug for long-running jobs, cutting-edge optimizations (Unified Memory, KV cache, kernels, speculative decoding), and benchmarking & production performance signal (MLPerf and automated alerting). You'll partner closely with NVIDIA and the open-source community while collaborating across Inference, Training, Orchestration (SUNK, Kueue, Kubeflow), Networking, Storage, and Product/Marketing to set the bar for performance and transparency. About the role: 1. Performance Tooling & Debug (Long Training Jobs; GPU↔Systems) Build and standardize CoreWeave's performance engineering toolkit to accelerate triage and root-cause analysis across GPU, network, storage, and training code with a special focus on GPU slowdowns and GPU SDCs (silent data corruptions) detection. • Unified observability: Integrate low-overhead profiling, tracing, and metrics (e.g., Nsight/CUPTI, DCGM/NVML, eBPF, NCCL telemetry) into a cohesive, Kubernetes-native experience spanning pods, nodes, and clusters., • Cross-layer correlation: Time-align GPU kernels, NCCL collectives, NIC/IB counters, storage IOPS/latency, and training logs to pinpoint bottlenecks across the stack., • SDC & correctness guardrails: Establish detection heuristics and workflows (redundancy checks, invariants, data/gradient sanity tests, ECC/firmware signals) for early SDC discovery and isolation., • Reproducible diagnostics: One-click capture bundles (configs, container images, HW/FW manifests, logs, traces) and runbooks to speed incident response and post-mortems., • Scale & safety: Bake performance gates into CI/CD and cluster admission, preventing regressions before they reach customers. 2. Global Performance Optimizations (Unified Memory, KV Cache, Kernels, Speculative Decoding, etc.) Drive a roadmap of cross-model and model-specific speedups for LLM training and inference. • KV Cache: Layouts, sharding, paging, and compression; optimized attention pathways (paged/persistent attention, Flash-style blocks) to unlock higher throughput and longer contexts., • Kernels: Fused and persistent kernels, launch configuration autotuning, compiler/toolchain upgrades, Tensor Core utilization, and memory-traffic reductions across precisions (BF16/FP8/INT8)., • Speculative decoding: Draft-model strategies, acceptance policies, and batching interplay to achieve lower latency and higher tokens/s without quality regressions., • Distributed performance: End-to-end throughput via NCCL topology tuning, multi-rail IB, NVLink/NVSwitch affinity, and data-loader/IO pipelines., • Upstream impact: Co-development with NVIDIA (CUDA/cuDNN/TensorRT-LLM/Triton, NCCL) and OSS (vLLM, PyTorch/DeepSpeed/Megatron-LM, ONNX Runtime, KServe) to land durable wins. 3. Benchmarking Systems & Production (MLPerf/Benchmarks + performance alerting) Deliver authoritative, auditable benchmarks and live service performance signals that customers can trust. • MLPerf ownership: Lead Training & Inference submissions—workload selection, cluster planning, audits, publication—and coordinate optimization tracks to achieve competitive results., • Internal benchmarking service: Operate a Kubernetes-native, repeatable service across SUNK (Slurm on Kubernetes), Kueue, and Kubeflow pipelines to measure p50/p95/p99 latency, jitter, tokens/s, TTFT, and cost per token/request across models, precisions, batch sizes, and GPU types., • Automated alerting: Detect and attribute HW/model/firmware changes that impact performance across LLM families, GPU SKUs, clusters, and regions; integrate with Prometheus/Grafana/OpenTelemetry and results warehouses for trend regression alarms., • Strategy & Roadmap: Define a multi-year performance strategy spanning tooling, optimizations, and benchmarks; prioritize models/workloads (LLMs, diffusion, vision, speech) and hardware tiers., • Organization Building: Hire, mentor, and develop performance engineers, kernel specialists., • Tooling & Automation: CI/CD for scheduled benchmarks and perf tests at fleet scale, creating data warehouse for performance analysis., • Cross-Functional Impact: Partner with Inference, Training, Orchestration, Networking, Storage, Solution architects, and Marketing. Who You Are:, • 12+ years in performance engineering for distributed systems/HPC/ML; 5+ years building & leading high-performing teams or orgs., • Demonstrated experience building large scale distributed systems that are highly available., • Demonstrated success driving end-to-end MLPerf (Training and/or Inference) or equivalent audited benchmarks at scale., • Deep understanding of GPU performance (CUDA, NCCL, RDMA, NVLink/PCIe, memory bandwidth), model-server stacks (Triton, vLLM, TensorRT-LLM, TorchServe), and distributed training (PyTorch FSDP/DeepSpeed/Megatron-LM). Why CoreWeave? At CoreWeave, we work hard, have fun, and move fast! We're in an exciting stage of hyper-growth that you will not want to miss out on. We're not afraid of a little chaos, and we're constantly learning. Our team cares deeply about how we build our product and how we work together, which is represented through our core values: • Be Curious at Your Core, • Act Like an Owner, • Empower Employees, • Deliver Best-in-Class Client Experiences The base salary range for this role is $206,000 to $303,000. The starting salary will be determined based on job-related knowledge, skills, experience, and market location. We strive for both market alignment and internal equity when determining compensation. In addition to base salary, our total rewards package includes a discretionary bonus, equity awards, and a comprehensive benefits program (all based on eligibility). What We Offer The range we've posted represents the typical compensation range for this role. To determine actual compensation, we review the market rate for each candidate which can include a variety of factors. These include qualifications, experience, interview performance, and location. In addition to a competitive salary, we offer a variety of benefits to support your needs, including: • Medical, dental, and vision insurance - 100% paid for by CoreWeave, • Company-paid Life Insurance, • Voluntary supplemental life insurance, • Short and long-term disability insurance, • Flexible Spending Account, • Health Savings Account, • Tuition Reimbursement, • Ability to Participate in Employee Stock Purchase Program (ESPP), • Mental Wellness Benefits through Spring Health, • Family-Forming support provided by Carrot, • Paid Parental Leave, • Flexible, full-service childcare support with Kinside, • 401(k) with a generous employer match, • Flexible PTO, • Catered lunch each day in our office and data center locations, • A casual work environment While we prioritize a hybrid work environment, remote work may be considered for candidates located more than 30 miles from an office, based on role requirements for specialized skill sets. New hires will be invited to attend onboarding at one of our hubs within their first month. Teams also gather quarterly to support collaboration California Consumer Privacy Act - California applicants only CoreWeave is an equal opportunity employer, committed to fostering an inclusive and supportive workplace. All qualified applicants and candidates will receive consideration for employment without regard to race, color, religion, sex, disability, age, sexual orientation, gender identity, national origin, veteran status, or genetic information. As part of this commitment and consistent with the Americans with Disabilities Act (ADA), CoreWeave will ensure that qualified applicants and candidates with disabilities are provided reasonable accommodations for the hiring process, unless such accommodation would cause an undue hardship. If reasonable accommodation is needed, please contact: . Export Control Compliance This position requires access to export controlled information. To conform to U.S. Government export regulations applicable to that information, applicant must either be (A) a U.S. person, defined as a (i) U.S. citizen or national, (ii) U.S. lawful permanent resident (green card holder), (iii) refugee under 8 U.S.C. § 1157, or (iv) asylee under 8 U.S.C. § 1158, (B) eligible to access the export controlled information without a required export authorization, or (C) eligible and reasonably likely to obtain the required export authorization from the applicable U.S. government agency. CoreWeave may, for legitimate business reasons, decline to pursue any export licensing process.