Loading…
Wednesday September 9, 2026 11:55 - 12:35 EDT
This session introduces a novel open source Domain-Specific Language (DSL) that enables developers to author high-performance GPU kernels directly in safe Rust. The DSL employs a tile-based programming model targeting the Tile IR open source MLIR dialect, allowing developers to focus on algorithms that decompose large tensor-based computations into smaller tile-based ones. By automating memory management and the utilization of specialized hardware like matrix accelerators, the DSL raises the GPU programming abstraction, compiling to high-performance binaries without the need for unsafe Rust.

To orchestrate workloads on modern systems, a tensor programming API enables the static composition of data movement and kernel launch operations into discrete device operations. These device operations are asynchronously submitted as tasks to multiple GPUs using a preferred async Rust runtime. Consequently, the async runtime serves as a unified scheduler of concurrent activities running on both the CPU and GPU.

Through an exploration of this system, attendees will gain a fundamental understanding of why data races occur in GPU code and see how they are fundamentally eliminated through the application of Rust’s ownership model. The session illustrates how to enforce memory safety within a macro-based DSL and extends those guarantees to asynchronous orchestration, providing a practical framework for building high-performance heterogeneous systems.
Speakers
avatar for Melih Elibol

Melih Elibol

Senior Research Scientist, NVIDIA
Melih Elibol is a Senior Research Scientist at NVIDIA, where he works on programming systems for GPU computing. He created cutile-rs, a safe Rust DSL for tile-based CUDA kernels. He holds a PhD from UC Berkeley, where he focused on distributed systems for machine learning.
Wednesday September 9, 2026 11:55 - 12:35 EDT
Palais des Congrès de Montréal

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

Share Modal

Share this link via

Or copy link