Workshops Archive

A Study In High Performance Portability Through Layered Abstractions


Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC

Authors: Stephen Jones (NVIDIA)

Abstract: In the past year NVIDA's CUDA programming model has grown to include several language extensions and DSLs, most notably with the addition of CUDA Tile, a tile-based parallel programming abstraction. We implemented CUDA Tile not just as a set of language features but as a series of compiler stages based around MLIR, allowing for much richer and more sophisticated code generation.

The effectiveness of this deeper compiler stack and the interaction between thread-level and tile-level parallelism quite honestly took us by surprise, and has provided remarkable insights into the mechanisms by which retargetable parallel abstractions provide excellent performance with a fraction the amount of code, requiring only auto-tuning with no code modification. NVIDIA's accelerated maths libraries are being converted to use CUDA Tile as a result.

We will share these insights and our analysis in this talk, along with additional conclusions from an investigation into how AI coding agents both benefit from and are impeded by Tile- and Thread-level abstractions when optimizing code.


Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing

Back to Full Workshop Archive Listing