Abstract
This research delves into the potential of Graphcore's IPU AI-optimized architecture in the realm of High-Performance Computing (HPC). Focusing on solving complex challenges like sparse systems and irregular applications, the dissertation contributes to HPC systems and application design, backed by rigorous empirical evaluation.The core objective is to evaluate the applicability of recent AI advancements to the field of HPC, with a specific focus on sparse and irregular applications. The thesis leverages empirical research, encompassing detailed experiments and analysis. This not only demonstrates how applications can seamlessly integrate with the AI accelerator but also highlights potential challenges and limitations associated with its use.In a world increasingly reliant on high-performance computing, this work's findings carry implications for both the scientific and technological communities. By bridging the gap between AI-specific architectures and HPC, the dissertation paves the way for more efficient problem-solving in diverse applications, offering insights for future advancements in this critical intersection of fields.
List of papers
Paper I: Luk Burchard, Xing Cai, Johannes Langguth. “iPUG for Multiple Graphcore IPUs: Optimizing Performance and Scalability of Parallel Breadth-First Search”. In: 2021 IEEE 28th International Conference on High Performance Computing, Data, and Analytics (HiPC). Vol. 28, (2021, December), pp. 162–171. DOI: 10.1109/HiPC53243.2021.00030. An accepted version of the article is included in the thesis. Also available at: https://doi.org/10.1109/HiPC53243.2021.00030 |
Paper II: Luk Burchard, Kristian Gregorius Hustad, Johannes Langguth, Xing Cai. “Enabling Unstructured-Mesh Computation on Massively Tiled AI-Processors:
An Example of Accelerating In-Silico Cardiac Simulation”. In: Frontiers in Physics. Vol. 11, (2023, March), pp. 105. DOI: 10.3389/fphy.2023.979699. An accepted version of the article is included in the thesis. Also available at: https://doi.org/10.3389/fphy.2023.979699 |
Paper III: Max Xiaohang Zhao, Luk Burchard, Daniel Thilo Schroeder, Johannes Langguth, Xing Cai. “iPuma: High-throughput Sequence Alignment for MIMD AI Accelerators”. The paper is not available in DUO awaiting publishing. |
Paper IV: Luk Burchard, Max Xiaohang Zhao, Johannes Langguth, Aydın Buluç, Giulia Guidi. “Space Efficient Sequence Alignment for SRAM-Based Computing: XDrop on the Graphcore IPU”. Accepted for publication in SC ’23: Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis. DOI: 10.1145/3581784.3607094. An accepted version of the article is included in the thesis. Also available at: https://doi.org/10.1145/3581784.3607094 |