CUDA Kernel Engineer — Optimize GPU Performance at Scale

Pragmatike; California (MO); On-site; USD 180,000 - 240,000; Full time

Benefits offered by this job

Health, Dental, and Vision

Sign-on bonus

401k

Job summary

Pragmatike is seeking an experienced GPU performance engineer to design and optimize CUDA kernels for NVIDIA GPUs in a fast-growing AI startup. You will optimize memory usage, profiling methods, and kernel launches to maximize throughput.

Responsibilities include challenging performance tuning, collaboration with AI and backend teams, and contributing to GPU architecture decisions. The role requires hands-on CUDA/C++ expertise and on-site work in SF.

Qualifications

• Proven track record building NVIDIA CUDA kernels from scratch.
• Strong ability to optimize kernels (tiling, occupancy, shared memory).
• Deep understanding of CUDA threads, warps, memory hierarchy, and warp divergence.

Responsibilities

• Design, implement, and optimize CUDA kernels for NVIDIA GPUs to maximize occupancy and throughput.
• Profile GPU workloads using Nsight Compute, Nsight Systems, nvprof, and CUDA-MEMCHECK.
• Analyze and eliminate performance bottlenecks like warp divergence and memory access issues.
• Improve memory pipelines (global/shared/L2/texture) and ensure memory coalescing.
• Collaborate with AI systems, model acceleration, and backend teams.
• Contribute to GPU architecture decisions and internal performance engineering practices.

Skills

CUDA kernel design

GPU performance optimization

CUDA memory management

C++ programming

Profiling and debugging

Tools

Nsight Compute

Nsight Systems

nvprof

CUDA-MEMCHECK

Job description

Pragmatike is seeking an experienced GPU performance engineer to design and optimize CUDA kernels for NVIDIA GPUs in a fast-growing AI startup. You will optimize memory usage, profiling methods, and kernel launches to maximize throughput.

Responsibilities include challenging performance tuning, collaboration with AI and backend teams, and contributing to GPU architecture decisions. The role requires hands-on CUDA/C++ expertise and on-site work in SF.