Afif Boudaoud

Zürich, Switzerland

I’m a PhD student at ETH Zürich working on the systems side of machine learning and scientific computing — how to make gradient-based methods fast enough to be useful on real workloads like climate models and large-scale Bayesian inference. My recent work includes ADELIA, the first AD-enabled INLA implementation, and DaCe AD, a unified automatic differentiation framework that’s on average over 92× faster than JAX on scientific computing patterns. I also contributed to the ICON weather model acceleration that won the 2025 ACM Gordon Bell Prize for Climate Modelling.

Before ETH, I was a research assistant at NYU Abu Dhabi working on learned compiler optimization, and a research intern at Mila on efficient deep learning. I did my Master’s at ESI in Algiers.

Research interests

High-performance automatic differentiation. Most AD frameworks force a choice between flexibility (JAX, PyTorch) and performance on real numerical workloads. My PhD is about building AD tools that don’t force that trade-off — especially for scientific computing patterns and structured problems like INLA where the underlying math has exploitable structure that generic AD throws away.

Compilers for scientific computing. Domain-specific code generation can match or beat hand-optimized code if the compiler understands the structure of the problem. I’m interested in how far that idea can be pushed — the ICON weather model work is one example, where structured transformations beat the production OpenACC version.

Learned approaches to code optimization. Cost models, scheduling, transformation search — there’s real value in learning these from data rather than hand-tuning heuristics, but the engineering of dataset generation and feature design matters enormously. This was what LOOPer and LOOPerSet were about.

selected publications

SC ’26
ADELIA: Automatic Differentiation for Efficient Laplace Inference Approximations

Afif Boudaoud, Lisa Gaedke-Merzhäuser, Alexandros Nikolaos Ziogas, and 6 more authors

arXiv preprint arXiv:2605.06392. Under review at SC ’26. , 2026

Abs arXiv Bib HTML

Spatio-temporal Bayesian inference drives environmental and health sciences using latent Gaussian models. INLA enables inference for these models at HPC scale but relies on derivative-based optimization over d hyperparameters. State-of-the-art INLA implementations approximate derivatives via central finite differences (2d+1 evaluations). ADELIA replaces this with reverse-mode automatic differentiation on a multi-GPU backward pass, achieving 4.2–7.9x per-gradient speedups and 5–8x energy savings over DALIA while enabling convergence where finite differences stall.
@article{boudaoud2026adelia, title = {ADELIA: Automatic Differentiation for Efficient Laplace Inference Approximations}, author = {Boudaoud, Afif and Gaedke-Merzh{\"a}user, Lisa and Ziogas, Alexandros Nikolaos and Maillou, Vincent and Calotoiu, Alexandru and Copik, Marcin and Rue, H{\aa}vard and Luisier, Mathieu and Hoefler, Torsten}, journal = {arXiv preprint arXiv:2605.06392}, year = {2026}, }
CLUSTER ’25
DaCe AD: Unifying High-Performance Automatic Differentiation for Machine Learning and Scientific Computing

Afif Boudaoud, Alexandru Calotoiu, Marcin Copik, and 1 more author

In Proceedings of the IEEE International Conference on Cluster Computing (CLUSTER), 2025

Abs arXiv Bib HTML

A general-purpose automatic differentiation engine that requires no code modifications and uses a novel ILP-based algorithm to balance storing vs. recomputing intermediates under a memory budget. Benchmarked on NPBench, it is on average over 92x faster than JAX on gradient computation for scientific computing patterns.
@inproceedings{boudaoud2025daceAD, title = {{DaCe AD}: Unifying High-Performance Automatic Differentiation for Machine Learning and Scientific Computing}, author = {Boudaoud, Afif and Calotoiu, Alexandru and Copik, Marcin and Hoefler, Torsten}, booktitle = {Proceedings of the IEEE International Conference on Cluster Computing (CLUSTER)}, year = {2025}, }

SC ’25

PerfDojo: Automated ML Library Generation for Heterogeneous Architectures

Andrei Ivanov, Siyuan Shen, Gioele Gottardo, and 5 more authors

In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC), 2025

DOI arXiv Bib HTML

@inproceedings{ivanov2025perfdojo,
  title = {{PerfDojo}: Automated ML Library Generation for Heterogeneous Architectures},
  author = {Ivanov, Andrei and Shen, Siyuan and Gottardo, Gioele and Chrapek, Marcin and Boudaoud, Afif and Schneider, Timo and Benini, Luca and Hoefler, Torsten},
  booktitle = {Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC)},
  year = {2025},
  doi = {10.1145/3712285.3759900},
}

PACT ’25

LOOPer: A Learned Automatic Code Optimizer for Polyhedral Compilers

Massinissa Merouani, Afif Boudaoud, Iheb Nassim Aouadj, and 7 more authors

In Proceedings of the International Conference on Parallel Architectures and Compilation Techniques (PACT), 2025

arXiv Bib HTML

@inproceedings{merouani2025looper,
  title = {{LOOPer}: A Learned Automatic Code Optimizer for Polyhedral Compilers},
  author = {Merouani, Massinissa and Boudaoud, Afif and Aouadj, Iheb Nassim and Tchoulak, Nassim and Bernou, Islem Kara and Benyamina, Hamza and Benbouzid-Si Tayeb, Fatima and Benatchba, Karima and Leather, Hugh and Baghdadi, Riyadh},
  booktitle = {Proceedings of the International Conference on Parallel Architectures and Compilation Techniques (PACT)},
  year = {2025},
}