Thomas Raoux
Member Of Technical Staff at OpenAI
Mountain View, California, United States
Summary
🤩
Rockstar🎓
Top SchoolThomas Raoux is a seasoned compiler engineer with 14 years of experience building high-performance GPU and accelerator toolchains, currently working on the Triton compiler at OpenAI. His background spans leading compiler efforts at Intel and Google, including the TPU LLVM backend and IREE CUDA optimizations, combining deep systems knowledge with practical performance engineering. He specializes in lowering, vectorization, shared memory management, and code-generation optimizations for matrix multiplication and GPU kernels. An active open-source contributor, he has made notable low-level improvements to Triton and IREE that tighten LLVM integration and boost kernel efficiency. Based in Mountain View, he brings a research-grade understanding of MLIR/LLVM internals together with production-focused delivery. Collected experience from embedded DSP compilers to modern ML compilers gives him an unusual fluency across decades of compiler and GPU evolution.
14 years of coding experience
15 years of employment as a software developer
Master, Computer science, Master, Computer science at Ecole centrale de Lyon
Github Skills (13)
compiler-optimization10
cuda10
matrix-multiplication10
triton10
code-generation10
shared-memory10
c-language10
cprogramming-language10
gpu10
vectorization10
llvm10
p89
machine-learning5
Programming languages (9)
Github contributions (5)
iree-org/iree
Jun 2020 - Jan 2023
A retargetable MLIR-based machine learning compiler and runtime toolkit.
Role in this project:
Back-end Developer & Performance Engineer
Back-end Developer & Performance EngineerContributions:1020 reviews, 409 commits, 871 PRs in 2 years 8 months
Contributions summary:Thomas's contributions focused on improving the performance of the IREE project's CUDA backend. They implemented a new pass to handle and optimize shared memory copy operations. The user also worked on enhancing the vectorization process, aiming to generate more efficient CUDA kernels through the use of vector operations, including a transformation that utilizes transfer read/write for vectorization. They further refined the approach to handling memory, particularly within the context of hardware specific needs to improve efficiency.
mlirspirvvulkantensorflowcompiler
triton-lang/triton
Jun 2023 - Apr 2025
Development repository for the Triton language and compiler
Role in this project:Back-end Developer
Back-end DeveloperContributions:1874 reviews, 839 PRs, 562 pushes in 1 year 9 months
Contributions summary:Thomas made contributions focused on enhancing the Triton language compiler, specifically related to the Triton GPU and LLVM backend. The commits involve implementing new features, such as the lowering of eviction policies for store operations, and extending existing functionalities. They also addressed critical issues, including fixing incorrect intermediate IR, and improving the performance of the code generation phase. The user's work primarily revolves around optimizing matrix multiplications and managing shared memory layouts.
compilerprogramming-languagecode-generationtriton
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.