LLM Inference Performance Engineer - GPU Kernel Optimizer

Intel; Folsom (CA); Hybrid; USD 171,000 - 315,000; Full time

Benefits offered by this job

Stock bonuses

Health benefits

Hybrid work model

Job summary

Intel is seeking a performance-focused AI Infrastructure Engineer to push LLM inference to its limits on Intel GPUs. You will optimize end-to-end inference pipelines, profile cross-stack bottlenecks, and develop high-performance kernels for attention, MoE, and quantization.

You will also upstream improvements into vLLM, SGLang, and PyTorch while shaping future GPU roadmaps. Work spans from kernel development to open-source contributions, with a hybrid work model and a competitive total

Qualifications

• 3+ years of relevant software engineering experience in GPU computing, AI systems, or HPC.
• Proficiency in modern C++ and Python; comfortable with complex systems-level code.

Responsibilities

• Drive Inference Performance: Own the end-to-end optimization pipeline for running state-of-the-art LLMs on Intel GPUs.
• Deep Stack Optimization: Profile, diagnose, and resolve cross-stack performance bottlenecks.
• Kernel Development and Integration: Design, write, and optimize custom high-performance kernels for attention, MoE, quantization, and operator fusions.
• Open Source Leadership: Upstream architectural improvements and hardware backends into open-source repos like vLLM, SGLang, PyTorch.
• Shape the Hardware Roadmap: Apply roofline analysis and profiling to decompose bottlenecks; partner with architecture and compiler teams.

Skills

C++

Python

GPU computing

Education

Bachelor's degree in relevant field

Master's degree (optional)

PhD (optional)

Tools

Triton

SYCL

CUDA/CUTLASS

Job description

Intel is seeking a performance-focused AI Infrastructure Engineer to push LLM inference to its limits on Intel GPUs. You will optimize end-to-end inference pipelines, profile cross-stack bottlenecks, and develop high-performance kernels for attention, MoE, and quantization.

You will also upstream improvements into vLLM, SGLang, and PyTorch while shaping future GPU roadmaps. Work spans from kernel development to open-source contributions, with a hybrid work model and a competitive total