Senior GPU Performance Software Engineer

Intel•Hillsboro, OR

3d•Hybrid

About The Position

The Software and AI (SAI) organization is seeking a highly skilled software engineer to contribute to the development and low-level optimization of oneDNN, a complex, cross-platform, open-source performance library that serves as the foundation for deep learning applications. This is a low-level software engineering and hardware-acceleration role focused on developing highly optimized math primitives, parallel algorithms, and GPU kernels that power industry-leading AI frameworks on Intel hardware.

Requirements

BSc, MSc, or PhD in Computer Science, Computer Engineering, Mathematics, Physics, or a highly technical related field
5+ years of professional software development experience with expert-level modern C++
2+ years of hands-on experience in programming and kernel optimization on GPUs (via SYCL/DPC++, OpenCL, CUDA, or HIP), or at least 5+ years of similar low-level performance optimization experience on CPUs
Strong foundations in computer architecture, cache hierarchies, memory subsystems, and parallel programming paradigms (e.g., multi-threading, SIMD/vectorization)

Nice To Haves

Experience developing high-performance math libraries (e.g., GEMM, convolution, reduction, or FFT kernels)
Hands-on experience with GPU assembly-level tuning or compiler optimization
Familiarity with parallel programming APIs such as OpenMP or oneTBB
Basic understanding of deep learning primitives (e.g., forward/backward passes) to understand how library code is utilized by upstream frameworks

Responsibilities

Develop high-performance GEMM, convolution, and attention kernels for AI workloads
Design scalable JIT and codegen infrastructure for GPU kernel generation
Implement fusion and memory-traffic optimizations to maximize hardware utilization
Optimize mixed-precision and quantized execution paths (e.g., BF16, FP16, INT8, FP8, FP4, etc.)
Build analytical and empirical performance models for kernel dispatch and tuning
Profile and eliminate performance bottlenecks across oneDNN GPU primitives and runtime paths
Co-design GPU primitives and kernel architectures for next-generation Intel GPUs
Partner with hardware and compiler teams to shape future accelerator capabilities and software stacks
Improve validation, benchmarking, and CI infrastructure for performance-critical GPU workloads