Minisymposium

As high performance computing (HPC) applications continue to push the boundaries of complexity and performance, leveraging AI-driven code generation has emerged as a powerful approach to streamline scientific software development processes.This talk explores the role of AI code generation models as a collaborative pair programmer in developing portable and scalable code with the SPH-EXA simulation framework.Specifically, we use AI-based code generation to assist development of SPH-EXA using CUDA, HIP, and SYCL to automate the code implementation and unit test generation and significantly reduce development effort while maintaining high standards of correctness and performance. Our work evaluates the dynamic interplay between human expertise and generative AI capabilities. We will assess productivity gains and discuss the practical challenges of ensuring performance portability and scalability across diverse hardware platforms. Through detailed performance analyses we will highlight both the benefits and the limitations of current AI-assisted approaches for scientific software development.This talk will provide insights into how AI-driven tools can serve as effective collaborative partners in HPC code development.

As the landscape of machine learning (ML) continues to evolve, the integration of generative AI has become a focal point for automating code generation. While it is perfectly suitable to generate text for humans such as the abstract you're reading, this approach often falls short in ensuring the correctness of the generated code, leading to potential pitfalls in robust ML applications. Recent cases exemplify this challenge, where the use of AI to accelerate coding processes resulted in faster but incorrect code, something unacceptable in scientific computing. In response to this critical need, we introduce TADASHI—a novel library designed to bridge the gap between speed and correctness in code transformations. TADASHI offers a user-friendly Python interface that can seamlessly integrate into existing ML scripts. It empowers developers by not only expediting code modifications but also enforcing rigorous correctness checks on these transformations. By ensuring that code remains valid and reliable, TADASHI enhances the stability of ML workflows and fosters confidence in automated processes. Join us as we delve into TADASHI's capabilities, showcasing its potential to revolutionize the code transformation landscape in ML, ensuring that efficiency does not compromise correctness.

Software performance evolves over time, yet traditional debugging and profiling remain reactive, costly, and disconnected from development workflows. Performance drift—gradual degradation in execution efficiency due to code modifications—often goes undetected until it causes significant slowdowns, forcing late-stage debugging and costly fixes. This talk presents a vision for AI/ML-driven proactive performance-drift detection, where models continuously monitor software evolution, identifying inefficiencies before they degrade execution. By combining static analysis (abstract syntax trees) with dynamic insights from nightly tests, this framework enables early detection of performance-impacting changes. Traditional ML approaches require full model retraining whenever code changes or new runtime data become available, making them impractical for fast-moving development cycles. Few-shot learning eliminates this overhead by allowing models to update incrementally with minimal new data. Attention-based representation learning further enhances interpretability by prioritizing performance-critical features, enabling more targeted interventions. This framework supports two key decision-making processes where (1) developers can receive automated feedback on whether a code change improves or degrades performance, enabling early intervention; (2) the insights can guide hardware configuration choices and runtime parameter tuning. This approach can be seamlessly integrated into CI/CD pipelines to achieve software that not only remains correct but also maintains efficiency as it evolves.

Modern GPUs, together with larger datasets, facilitate the exponential growth and adoption of deep learning models. The training and deployment of deep neural networks in widely used large-scale data centers, on the other hand, exhibit low GPU hardware utilization, barely reaching 50%, as shown by studies done on Microsoft and Alibaba clusters. This is a waste of hardware resources, especially for expensive GPUs, and contributes to the unsustainable carbon footprint of AI. In fact, less than 15% of large-language model training can lead to a carbon footprint equivalent to the average yearly energy consumption of a US household.This talk will discuss the reasons, challenges and opportunities for designing energy-efficient computing infrastructures as well as its practical applications.