[Sponsor Invited Talks'25] Programming Language Design for GPU Systems

Michel Steuwer discusses the challenges of programming GPU systems and introduces Descend, a safe GPU systems programming language.

Programming Language Design for GPU Systems (Video, Sponsorship Invited Talks 2025) Michel Steuwer (Technische Universität Berlin) Abstract: Programming GPU systems poses special challenges. Mixing GPU and CPU hardware architectures to achieve high performance is a demanding task even for advanced programmers. One might think that programming languages are a key tool for empowering programmers to confidently write correct and fast software for GPU systems. I agree! However, maybe surprisingly few programming languages for heterogeneous and high-performance systems have been proposed. Most common programmers still write CUDA paired with C++ or possibly Python, the same languages from when GPUs opened up for general computations almost two decades ago. In this talk, I will discuss opportunities for programming language design to improve the productivity of developers and help them to write better software exploiting the performance potential of modern GPU systems. I will showcase our experience of developing Descend, a Rust-inspired programming language for heterogeneous systems with GPUs, and discuss advantages but also current limitations of advanced type systems for developing high-performance software. I will also look ahead and point out future directions where programming language design could play a vital role in making the development of software for heterogeneous systems more robust and efficient. Presentation at the Sponsorship Invited Talks 2025 forum, Oct 18, 2025, conf.researchr.org/home/icfp-splash-2025/industry-forum-2025 Sponsored by ACM SIGPLAN.