Software Engineer, AI Kernels & Performance Optimization — MTIA Software

Meta; Bellevue (WA); Hybrid; USD 180,000 - 260,000; Full time

Job summary

Meta is seeking an experienced kernel and performance engineer to own the performance of AI workloads at scale. You will work on roofline-level kernels, from fused attention to numerics decisions, and read hardware specs and RTL-adjacent docs to diagnose issues.

You will contribute to kernel frameworks and pre-silicon bring-up, shaping the next generation of AI accelerators. You will collaborate with teams across Meta Infrastructure and the AI & Compute Foundation to deliver high-performance,

Qualifications

• 6+ years of professional experience in high‑performance computing, accelerator kernel development, compiler backends, or systems performance engineering
• Proficiency in C++ and Python, including low-level systems programming, templates and generic programming
• Experience writing and optimizing kernels for parallel architectures such as GPU, TPU or other AI ASICs
• Familiarity with computer architecture concepts: memory hierarchies, bandwidth, latency hiding, vectorization
• Experience with MLIR, LLVM, XLA, Halide or related compiler/codegen tech
• Pre‑silicon software development and hardware/software co‑design experience

Responsibilities

• Own roofline-level kernels (GEMM, attention variants) and scalar/fused operations
• Develop and optimize kernel libraries and frameworks for rapid performance tuning
• Read hardware specs and RTL-adjacent docs to identify software bottlenecks
• Contribute to pre-silicon bring-up and validation on simulators and emulators
• Mentor engineers and guide technical direction across teams

Skills

High-performance computing

Accelerator kernel development

Compiler backends

Performance engineering

C++ programming

Python programming

GPU/AI accelerator experience

Pre-silicon software development

Education

Bachelor's degree in Computer Science / Computer Engineering / related field

Tools

CUDA

LLVM/MLIR

Triton

CUTLASS

PyTorch

Job description

Meta designs and deploys its own AI systems. MTIA — the Meta Training and Inference Accelerator — is Meta's family of in-house AI accelerator ASICs, running recommendation and ranking workloads in production across Meta's data centres today and expanding into generative AI inference and training as successive silicon generations land (see https://bit.ly/metamtia).

The MTIA Software team is part of the AI & Compute Foundation (ACF) organization within Meta Infrastructure. Because the hardware is ours, the software is ours too: we build the entire stack a chip vendor would normally supply — compiler and LLVM toolchain, runtime, kernel authoring frameworks and libraries, developer tooling, and deep PyTorch integration — and we co-design it with the silicon teams generation over generation.

Within that stack, the AI kernel and optimization software development team drives the layer where architecture meets arithmetic. Our mission is performance *and* programmability at scale: hit roofline enablements on the workloads that matter, and make kernel authoring accessible enough that the whole organization can close coverage gaps without funneling every problem through a handful of experts. We do this by shipping high-performance kernel libraries with broad PyTorch operator coverage, by building the C++ and Python kernel authoring frameworks and DSL surfaces that others build on, and by writing production kernels against new architectures long before first silicon — turning hardware proposals into measured roofline evidence while the design can still change.

We are hiring an experienced kernel and performance engineer to take on this work at a senior level. You will own the performance of workloads that serve billions of people, from the innermost loop of a fused attention kernel to the numerics decisions that determine whether a model converges. You will read hardware specifications and RTL‑adjacent documentation as easily as you read code, and you will be expected to say clearly when the hardware — not the software — is the problem. Your findings will change what gets built next.

What you’ll work on

• Roofline-level kernels. GEMM and attention variants, normalization, collectives, elementwise and reduction fusions, sparse and quantized paths — implemented against novel architectural features (matrix engines, on‑chip reduction fabrics, software‑managed memory hierarchies) and tuned until the remaining gap to the machine's limit is explainable in a sentence.
• Numerics under precision constraints. Low-precision formats (FP8, MX‑style block‑scaled types, integer quantization) where the difference between a correct scale choice and a plausible one is several dB of signal, and where the fix has to work on silicon that has already been taped out.
• Kernel authoring frameworks. Templateized, composable C++ kernel SDKs in the spirit of CUTLASS, Python DSLs in the spirit of Triton and CuTe, and the compiler‑facing interfaces that let automated codegen reach performance that used to require a specialist.
• Pre‑silicon and bring‑up. Kernels on simulators and emulators, validating architectural features and rooflines before tapeout, then first‑light bring‑up on real parts.
• Software mitigations for hardware reality. Every chip ships with something you wish were different. Finding the workaround that recovers most of the lost performance — and generalizing it so nobody rediscovers it — is core to the job.
• Bachelor's degree in Computer Science, Computer Engineering, relevant technical field, or equivalent practical experienceBachelor's degree in Computer Science, Computer Engineering, a related technical field, or equivalent practical experience6+ years of professional experience in high‑performance computing, accelerator kernel development, compiler backends, or systems performance engineeringProficiency in C++ and Python, including low-level systems programming, templates and generic programming, and comfort reading and writing performance‑critical codeDemonstrated experience writing and optimizing kernels for a parallel architecture — GPU (CUDA, ROCm/HIP, SYCL/OpenCL), TPU or other AI ASICs, or SIMD/vector CPU targetsWorking knowledge of computer architecture as it applies to performance: memory hierarchies and bandwidth, latency hiding, occupancy and scheduling, vectorization, and synchronizationA rigorous, measurement‑driven approach to performance: the ability to build a roofline or analytical model, profile against it, and explain the residual gap Experience mentoring engineers and setting technical direction across teamsExperience adhering to and implementing responsible, ethical AI practices (e.g., risk assessment, bias mitigation, quality and accuracy reviews)Experience with distributed execution and collective communication (NCCL/RCCL‑class primitives, tensor and expert parallelism, overlapping communication with compute)Experience with low-precision numerics and quantization — FP8/E4M3/E5M2, MX and other block‑scaled formats, INT8/INT4 — including error analysis and calibrationExperience with compiler and codegen technologies relevant to kernels: MLIR, LLVM, TVM, XLA, Halide, or polyhedral scheduling8+ years of experience in accelerator software, HPC, or ML systems performance (or equivalent with an advanced degree)Deep familiarity with transformer and attention kernel design: FlashAttention‑class algorithms, KV‑cache management, paged and chunked attention, linear and state‑space attention variants, MoE routing and expert dispatchTrack record of open‑source contribution in the kernel, compiler, or ML systems ecosystemExperience with pre‑silicon software development — architectural simulators, FPGA emulation, performance modeling — and with hardware/software co‑design cyclesFamiliarity with ML framework internals: PyTorch dispatch and eager execution, torch.compile / Inductor, custom operator integration, and inference serving stacks such as vLLM or SGLangExperience building or contributing to high‑performance kernel libraries or frameworks — CUTLASS, cuBLAS, cuDNN, CUTE, Triton, Helion, ThunderKittens, oneDNN, Composable Kernel, or comparable internal equivalentsDemonstrated ongoing AI skill development (e.g., prompt/context engineering, agent orchestration) and staying current with emerging AI technologiesDemonstrated ability to integrate AI tools to optimize/redesign workflows and drive measurable impact (e.g., efficiency gains, quality improvements)