Aroun Demeure
Member Of Technical Staff at Magic
San Francisco, California, United States
Summary
🤩
Rockstar🎓
Top SchoolAroun Demeure is a GPU and AI-focused engineer based in San Francisco with deep expertise in low-level C/C++/CUDA performance engineering and hardware-software co-design. He has driven architecture and performance work at Imagination Technologies and Apple, inventing microarchitectural optimizations across GPU pipelines and influencing AI tensor core trade-offs. At Imagination he led major projects (E-Series/D-Series) and created long-lived performance tools and compiler optimizations that materially reduced instruction counts and power. More recently he contributed high-impact backend and CUDA optimizations to the popular karpathy/llm.c repository, improving memory management and fused kernels for faster LLM training. Fluent in French and English, with an ongoing BSc in Mathematics, he blends rigorous analytical thinking with hands-on systems tuning across hardware, drivers, and ML training stacks. Outside work he pursues interests spanning visual programming, evolutionary biology, and efficient network programming, reflecting a broad curiosity beyond conventional GPU engineering.
2 years of coding experience
6 years of employment as a software developer
BSc Mathematics (Part-Time, Ongoing), BSc Mathematics (Part-Time, Ongoing) at The Open UniversityHigh School, High School at Athénée Royal Crommelynck
Github Skills (7)
cuda10
cublas10
performance-optimization10
cprogramming-language9
c-language9
bfd9
machine-learning8
Programming languages (2)
Github contributions (5)
karpathy/llm.c
Apr 2024 - Oct 2024
LLM training in simple, raw C/CUDA
Role in this project:
Back-end Developer & Performance Engineer
Back-end Developer & Performance EngineerContributions:23 reviews, 39 PRs, 67 comments in 5 months
Contributions summary:Aroun significantly improved the performance of the LLM training process. Their contributions focused on optimizing CUDA code, including the use of `cudaHostMalloc` for memory management and the implementation of a new softmax kernel optimized for large vocabulary sizes. Furthermore, they made changes to leverage cuBLAS and cuBLASLt for faster matrix multiplication operations and also introduced a novel fused classifier with softmax for efficiency, resulting in improved training speed. The code changes involved the modification of the core training loop.
ademeure/llm.c
Apr 2024 - Nov 2024
LLM training in simple, raw C/CUDA
Contributions:2 reviews, 2 PRs, 243 pushes in 7 months
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.