Senior AI Kernel & Performance Engineer - Pre-Silicon

Meta; New York (NY); On-site; USD 184,000 - 257,000; Full time

Job summary

Meta is hiring an experienced kernel and performance engineer for MTIA Software to own the performance of workloads across MTIA accelerators. You will read hardware specs, optimize kernels (GEMM, attention variants), and collaborate with compiler, runtime, and framework teams to push production models toward higher throughput and efficiency.

You will design templates and libraries for kernel authoring, work on pre-silicon bring-up, and develop software mitigations that recover performance lost

Qualifications

• Bachelor's degree in CS/CE or related field and practical experience.
• 6+ years in HPC, accelerator kernel development, or systems performance.
• Proficiency in C++ and Python for performance-critical code.
• Experience writing/optimizing kernels for parallel architectures (GPU/ASIC).
• Strong knowledge of memory hierarchies, latency hiding, and roofline analysis.

Responsibilities

• Design, implement, and optimize high-performance compute kernels for MTIA accelerators.
• Profile and root-cause performance across full stack: kernel, compiler, runtime, hardware.
• Build kernel authoring frameworks and libraries for broad PyTorch operator coverage.
• Work with silicon teams on hardware/software co-design and pre-silicon bring-up.

Skills

C++

Python

Kernel development

Education

Bachelor's degree in Computer Science, Computer Engineering, or related field

Tools

CUDA

LLVM

PyTorch

Job description

Meta is hiring an experienced kernel and performance engineer for MTIA Software to own the performance of workloads across MTIA accelerators. You will read hardware specs, optimize kernels (GEMM, attention variants), and collaborate with compiler, runtime, and framework teams to push production models toward higher throughput and efficiency.

You will design templates and libraries for kernel authoring, work on pre-silicon bring-up, and develop software mitigations that recover performance lost