Kernel Developer

About the Role

We are looking for a C/C++ engineer to design and optimize high-performance compute kernels for AI accelerators.

This is not Linux kernel development, and it does not involve drivers, operating systems, or kernel modules.
In this context, “kernels” refer to user-space compute kernels for tensor operations used in deep learning workloads.

You will work at the lowest level of AI performance engineering — where software meets specialized hardware.

Responsibilities

• Design and implement high-performance compute (operator) kernels in C/C++

• Develop core tensor operations

• Optimize performance for AI accelerators (latency, throughput, and efficiency)

• Apply low-level optimization techniques

• Profile, benchmark, and tune kernels to eliminate performance bottlenecks

• Contribute to internal libraries and runtime systems for AI workloads

Requirements

• Strong proficiency in C/C++

• Experience with performance-critical software development

• Strong understanding of low-level optimization techniques

• Understanding of CPU/GPU or accelerator architecture fundamentals

• Ability to analyze and debug complex systems

• Experience working with large, complex codebases

• Strong communication and teamwork skills

Nice to Have

• Experience with GPU kernel programming (CUDA / ROCm / OpenCL)

• Experience with Triton or similar kernel programming frameworks

• Knowledge of instruction set architectures (ISA)

• Familiarity with compiler technologies (e.g., LLVM-based stacks)

• Experience with distributed communication frameworks (NCCL, MPI, libfabric)

• Understanding of deep learning models

What We Offer

• Highly competitive salary, employment contract (Umowa o Pracę), and a comprehensive benefits package, including Medicover healthcare coverage.

• Work on the performance-critical compute layer for next-generation AI accelerators

• Direct impact on deep learning model efficiency and latency

• Collaboration with experts in hardware, compilers, and systems

• Challenging low-level performance engineering problems at the hardware–software boundary