GPU Kernel Engineer-High-Performance AI Primitives (Hybrid)

Talanto; Hillsboro, Northern (OR, KY); Hybrid; USD 195,000 - 276,000; Full time

Benefits offered by this job

Hybrid work model

Stock bonuses

Health insurance

Retirement plan

Paid vacation

Job summary

Talanto is seeking a Senior GPU Performance Software Engineer to contribute to oneDNN development and low-level optimization for AI workloads on GPUs. This is a low-level role focusing on high-performance math primitives, kernel design, and codegen for next-generation accelerators.

You will collaborate with hardware teams and open-source communities, optimize FP16/BF16 paths, and shape performance across OpenCL/CUDA/HIP stacks while supporting CI and validation pipelines.

Qualifications

• Education: BSc, MSc, or PhD in Computer Science, Computer Engineering, Mathematics, Physics, or related fields.
• 5+ years of professional software development with expert-level C++.
• 2+ years of GPU kernel optimization or 5+ years low-level optimization on CPUs.
• Strong foundations in computer architecture, memory subsystems and parallel programming.

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.
• Profile and optimize mixed-precision and quantized execution paths (BF16, FP16, INT8, FP8, FP4).

Skills

Modern C++

GPU kernel optimization

Parallel programming

Performance tuning

Education

BSc/MSc/PhD in CS or related

Tools

SYCL/DPC++

OpenCL/CUDA/HIP

Job description

Talanto is seeking a Senior GPU Performance Software Engineer to contribute to oneDNN development and low-level optimization for AI workloads on GPUs. This is a low-level role focusing on high-performance math primitives, kernel design, and codegen for next-generation accelerators.

You will collaborate with hardware teams and open-source communities, optimize FP16/BF16 paths, and shape performance across OpenCL/CUDA/HIP stacks while supporting CI and validation pipelines.