GPU Kernel Optimization Engineer for AI Training

Obsidian; Chicago (IL); On-site; USD 83,000 - 165,000; Full time

Job summary

Mercor is seeking GPU kernel optimization experts to contribute to a project with a leading AI lab. This contract-based opportunity targets freelancers with strong C++ skills, practical GPU programming experience, and the ability to improve kernel performance using profiler-guided analysis.

You will evaluate, optimize, and reason about GPU kernels across modern hardware environments, writing C++17, Python, and CUDA/HIP code, and documenting optimization decisions clearly.

Qualifications

• Strong C++ skills and GPU programming experience.
• Experience with profiler-guided optimization and kernel performance analysis.
• Familiarity with NVIDIA/AMD/Qualcomm GPU architectures is a plus.

Responsibilities

• Analyze and optimize GPU kernels for performance, efficiency, and hardware utilization.
• Use profiler metrics such as L2 cache hit rate, L2 throughput, occupancy to guide kernel improvements.
• Review GPU kernel implementations and identify bottlenecks without requiring extensive background in the underlying algorithms.
• Write, modify, and reason about C++17, Python, and GPU programming code.
• Apply CUDA, HIP, shader programming, or related kernel programming expertise to improve performance outcomes.
• Document optimization decisions clearly, including when specific profiler metrics are or are not useful.

Skills

C++17

GPU programming

Profiling skills

Python

Tools

CUDA

HIP

Slang

HLSL

GLSL

PTX

Tensor cores

NSight Compute

CUDA C++ Core Libraries

Job description

Mercor is seeking GPU kernel optimization experts to contribute to a project with a leading AI lab. This contract-based opportunity targets freelancers with strong C++ skills, practical GPU programming experience, and the ability to improve kernel performance using profiler-guided analysis.

You will evaluate, optimize, and reason about GPU kernels across modern hardware environments, writing C++17, Python, and CUDA/HIP code, and documenting optimization decisions clearly.