Senior GPU Kernel Engineer: Inference Throughput

CoreWeave; Bellevue (WA); On-site; USD 182,000 - 242,000; Full time

Benefits offered by this job

Medical, dental, and vision insurance

Company-paid Life Insurance

401(k) with employer match

Tuition Reimbursement

Flexible PTO

Catered lunch in office

Job summary

CoreWeave is seeking a Senior Engineer for its Benchmarking & Performance team to own kernel authoring and optimization for LLM inference. You will write and tune CUDA kernels, optimize tensor cores, and push end-to-end latency down while maintaining accuracy.

You will lead benchmarking workflows (MLPerf), mentor engineers, and collaborate with cross-functional partners. Experience with CUDA, C++, Python, and GPU architectures is required; familiarity with vLLM, TensorRT-LLM, llm-d, SGLang is a

Qualifications

• 5+ years of experience in HPC, GPU software, or performance-critical systems.
• Hands-on CUDA experience with written/optimized kernels and CUDA memory model.
• Deep understanding of GPU architecture, tensor cores, warp occupancy, memory hierarchy.

Responsibilities

• Author, profile, and optimize CUDA kernels for LLM inference on the critical path.
• Tune tensor cores, occupancy, memory coalescing, shared memory, and data movement overlap.
• Benchmark end-to-end latency/throughput; maintain reproducible MLPerf workflows and runbooks.

Skills

High-performance computing

CUDA experience

C++

Python

GPU architecture

Model-serving stacks

Cross-functional collaboration

Tools

Nsight Compute

Nsight Systems

NCCL

Kubernetes

Job description

CoreWeave is seeking a Senior Engineer for its Benchmarking & Performance team to own kernel authoring and optimization for LLM inference. You will write and tune CUDA kernels, optimize tensor cores, and push end-to-end latency down while maintaining accuracy.

You will lead benchmarking workflows (MLPerf), mentor engineers, and collaborate with cross-functional partners. Experience with CUDA, C++, Python, and GPU architectures is required; familiarity with vLLM, TensorRT-LLM, llm-d, SGLang is a