ICSAn Open-Source-First Approach: Multi-Level Compiler Design for AI and HPC
Writing compilers yourself (or understanding them in great detail) has become mission-critical for performance engineers in HPC and AI. HPC engineers still rely on high-performance C++ or Fortran compilers, OpenMP and MPI, and CUDA to program their applications. Yet, highly specialized hardware accelerators from Tenstorrent, Cerebras, and Nextsilicon, as well as higher-level programming models from NVIDIA, have widened the programming surface one needs to master to attain peak performance. While a plethora of new programming models are used to increase accessibility, designing a custom domain-specific compiler stack is surprisingly often an impactful option. In the following talk, I share my experience building multi-level compilers, from HPC to AI, within the MLIR ecosystem, covering MLIR’s GPU dialect and its use for scientific computing, MLIR’s first-class MPI support, our work on targeting Cerbera’s Waverscale engine, and our most recent effort of opening the LLVM backend at the ISA-level using MLIR – often using Python to accelerate development. I then discuss how our open-source-first research approach enabled us to deliver practically informed research that scales to real workloads, while contributing ideas and code back to the wider open-source ecosystem and helping to make compiler design a powerful tool for HPC and beyond.
Cite
@inproceedings{grosser2026opensource,
author = {Grosser, Tobias},
title = {{An Open-Source-First Approach: Multi-Level Compiler Design for AI and HPC}},
booktitle = {Proceedings of the 40th ACM International Conference on Supercomputing},
series = {ICS},
year = {2026},
month = {7},
publisher = {ACM},
doi = {10.1145/3797905.3813790},
url = {https://doi.org/10.1145/3797905.3813790}
}