Cross-Hardware Optimization Theory for Conflicting ML Workloads

Furutanpey et al. reveal that vendor-specific compilers (e.g., TVM, XLA) produce wildly divergent results for the same ML model on different hardware. This idea proposes a theory of optimization transferability: a formal framework to predict when an optimization (e.g., operator fusion) will generalize across architectures. By analyzing patterns in conflicting results (e.g., why TVM outperforms TensorRT on GPUs but not TPUs), we could derive hardware-agnostic optimization principles. For instance, fusions that increase arithmetic intensity might benefit GPUs but hurt accelerators with limited memory bandwidth (like AWS Trainium). This extends Milepost GCC’s feature-based predictions to ML-specific workloads, using insights from Liu et al.’s metric learning to prune ineffective optimizations. The outcome: a "universal optimizer" that avoids hardware-specific pitfalls—resolving the "contradictory results" problem plaguing current research.

References:

1. Automatic Selection of Compiler Optimizations by Machine Learning. Melih Peker, Özcan Özturk, Süleyman Yildirim, Mahiye Uluyagmur Öztürk (2023). Signal Processing and Communications Applications Conference.
2. Optimizing Machine Learning Operators and Models for Specific Hardware Using Apache-TVM. Kausthub Thekke Madathil, Abhinav Dugar, Nagamma Patil, Unnikrishnan Cheramangalath (2023). International Conference on Computing Communication and Networking Technologies.
3. AWS Trainium: The Journey for Designing and Optimization Full Stack ML Hardware. Nafea Bshara (2024). International Conference on Architectural Support for Programming Languages and Operating Systems.
4. Leveraging Neural Graph Compilers in Machine Learning Research for Edge-Cloud Systems. Alireza Furutanpey, Carmen Walser, Philipp Raith, P. Frangoudis, S. Dustdar (2025). arXiv.org.
5. Milepost GCC: Machine Learning Enabled Self-tuning Compiler. G. Fursin, Yuriy Kashnikov, Abdul Wahid Memon, Z. Chamski, O. Temam, Mircea Namolaru, E. Yom-Tov, Bilha Mendelson, A. Zaks, Eric Courtois, F. Bodin, Phil Barnard, Elton Ashton, Edwin V. Bonilla, John Thomson, Christopher K. I. Williams, M. O’Boyle (2011). International journal of parallel programming.
6. Iterative Compilation Optimization Based on Metric Learning and Collaborative Filtering. Hongzhi Liu, Jie Luo, Ying Li, Zhonghai Wu (2021). ACM Transactions on Architecture and Code Optimization (TACO).