Senior AI Performance Engineer - High-Performance Kernels

Openkrill; United States; Hybrid; USD 180,000 - 270,000; Full time

Benefits offered by this job

Healthcare coverage

RSU grants

Paid time off

Job summary

Modular is seeking a senior AI Performance Engineer to design, implement, and tune high‑performance attention kernels for new accelerator architectures. You will work on GPUs, NPUs, ASICs, and other accelerators, contributing to Modular’s Attention framework with reusable abstractions and hardware‑specific optimizations.

You will mentor teammates, write design docs, and collaborate with hardware partners to deliver production‑ready workloads and performant tokens to end cloud users.

Qualifications

• 5+ years of relevant experience in HPC, AI kernel development, or accelerator programming.
• Proven ability to write and optimize production-quality GPU/accelerator kernels.
• Strong understanding of attention algorithms and their performance tradeoffs.

Responsibilities

• Design and implement high-performance attention kernels for various attention variants.
• Build and optimize general AI kernels (matmul, softmax, normalization, activation, embedding).
• Contribute to Attention framework design for portability and hardware optimization.
• Profile workloads, identify bottlenecks, and tune against hardware limits and baselines.

Skills

GPU/accelerator kernels

Performance profiling

Attention mechanisms

Heterogeneous architectures

CUDA/Triton/OpenCL

Kernel optimization

Education

Advanced degree in CS/EE

Tools

CuTe/CUTLASS

Triton

MLIR/LLVM

PyTorch/TensorRT

Job description

Modular is seeking a senior AI Performance Engineer to design, implement, and tune high‑performance attention kernels for new accelerator architectures. You will work on GPUs, NPUs, ASICs, and other accelerators, contributing to Modular’s Attention framework with reusable abstractions and hardware‑specific optimizations.

You will mentor teammates, write design docs, and collaborate with hardware partners to deliver production‑ready workloads and performant tokens to end cloud users.