Senior AI Kernel & Performance Engineer (MTIA)

Meta; Menlo Park (CA); On-site; USD 154,000 - 217,000; Full time

Job summary

Meta is seeking an experienced kernel and performance engineer for MTIA Software to own accelerator workloads, from fused attention kernels to convergence decisions. You will read hardware specs and RTL-adjacent docs, and determine whether the hardware or software is the bottleneck.

This hands-on role covers kernel libraries, C++/Python tooling, and pre-silicon bring-up, with broad impact on PyTorch operators and production workloads.

Qualifications

• Bachelor's degree in CS/CE or related field.
• Experience writing and optimizing kernels for parallel architectures (GPU/ASIC).
• Strong reading of hardware specs and RTL-adjacent docs.
• Proven ability to measure and close roofline gaps.

Responsibilities

• Design, implement, and optimize high-performance compute and communication kernels for MTIA accelerators.
• Profile and root-cause performance across stacks to drive fixes.
• Build and extend kernel authoring frameworks and libraries.
• Deliver and maintain broad kernel coverage for PyTorch operators across workloads.
• Partner with silicon teams on hardware/software co-design and roofline validation.
• Mentor engineers on performance methodology and accelerator programming.

Skills

C++ & Python

Kernel performance engineering

GPU/AI accelerator experience

Performance profiling

Architecture awareness

Debugging performance issues

RTL/architecture reading

Education

Bachelor's degree in CS/CE or related field

Tools

CUTLASS

Triton / CuTe

LLVM / MLIR / XLA

PyTorch internals

NCCL / RCCL

Job description

Meta is seeking an experienced kernel and performance engineer for MTIA Software to own accelerator workloads, from fused attention kernels to convergence decisions. You will read hardware specs and RTL-adjacent docs, and determine whether the hardware or software is the bottleneck.

This hands-on role covers kernel libraries, C++/Python tooling, and pre-silicon bring-up, with broad impact on PyTorch operators and production workloads.